Scientific Accomplishments and Contributions
EMBYR wildfire model
From 1991-1993 I developed a spatially-explicit grid-based forest fire model, EMBYR (Ecological Model for Burning the Yellowstone Region), for simulating wildfire in Yellowstone National Park, USA. EMBYR is a probabilistic wildfire simulation that predicts potential burn patterns of large fires relative to variations in fuel types and weather patterns in an area. Ignitions can occur at random points or specific locations, and ignitions from firebrands can be simulated relative to fuel type. EMBYR requires a GIS layer of fuel types based upon age classes and species composition. Fire spread probabilities are specified for three possible fuel moisture conditions; wet, intermediate, or dry. Probabilities are then adjusted using one of three wind speed categories and one of eight wind directions. The output from EMBYR indicates the final burn pattern of one or more potential landscape-scale fires, allowing impacts from future fires to be estimated. Each run is different; many stochastic runs of the same wildfire event produces a “probability density cloud” that shows the full statistical range of behavior of that fire. The spatial resolution is 50m. EMBYR is easily parameterized, fast and efficient, and shows interactions between landscape pattern and process. EMBYR development was sponsored by the National Science Foundation (NSF) under "Causes and Consequences of Large-Scale Fires" (web page describing early results here, source code available here, see Publication #36, describing the EMBYR model and its behavior, with 306 citations).
In 1994, Bob Gardner (ORNL staff) and I used EMBYR to simulate the wildfire regime over the next millennium for the Greater Yellowstone Ecosystem under three alternative, synthetic, fractally-generated climate scenarios: dry, moderate, and wet. Fuel growth and tree succession under each scenario were simulated by a Markov transition probability model. We performed 10 replications of 1000 years into the future for each of the three climate scenarios (see Publication #17, with 131 citations). Total area burned each year was nearly constant, regardless of the climate; only the average size and intensity of the wildfires changed.
In a series of three papers (Agricultural and Forest Meteorology 1998, Ecological Modeling 2004, and Landscape Ecology 2006), Bob Keane of USDA Forest Service, Missoula Fire Lab, and G.J. Cary of the Canadian Forest Service independently compared and evaluated the sensitivity of the EMBYR model with four other wildfire simulation models (FIRESCAPE, LANDSUM, SEM-LAND, and LAMOS(DS)). EMBYR was found to be one of the most sensitive models to landscape fuel patterns, in terms of changes in area burned.
In 2004, Town Peterson and I used a modified version of EMBYR to predict the spread of the invasive Asiatic longhorned beetle, Anoplophora glabripennis, across its suitable range within the United States, as modeled by the GARP (Genetic Algorithm for Ruleset Prediction) niche model. EMBYR provided an excellent parallel to species dispersal; fires spread via ignition of adjacent areas, and also through longer-distance dispersal by means of ‘firebrands’ — similar to the ways that an invasive species spreads across a landscape. Our parameterization of EMBYR was only intended to assess the spatial pattern of invading populations, not the actual rates of spread. We initiated the EMBYR spread model at 32 known points of warehouse or tree infestations in North America (see Publication # 51, with 50 citations, acc. To BioOne).
Fractal Realizer and MapCurves
In 1997, I was generating synthetic fractal maps of resource distributions for testing models of foraging theories (see Publication #19, 60 citations). Soon after realizing this general need, I devised and programmed the Fractal Landscape Realizer, which generates synthetic landscape maps to user specifications. The alternative landscape realizations are not identical to the actual maps after which they are patterned, but are similar statistically (i.e., the areas and fractal patterns of each category are replicated). A fractal or self-similar pattern generator is used to provide a spatial probability surface for each category in the synthetic map. The Fractal Realizer preserves the fractal patterns of all the categories in the resulting synthetic landscape. Each synthetic landscape is one realization from among an infinite ensemble of possible fractal landscape map combinations. One can use the Fractal Landscape Realizer to simulate an actual example landscape on-the-fly at http://www.geobabble.org/cgi-bin/realizer/turing-map?3. Click reload to generate another custom synthetic landscape. Every fractal map realization is new and different.
The Fractal Realizer is useful as a generator of “neutral models” against which to test for the presence of natural spatial patterns. The Fractal Realizer generates null models using well-defined structuring processes which are under the users' control. Replicated landscape maps generated using the Fractal Realizer all possess statistical properties that are similar to a particular empirical landscape, and can provide a baseline upon which to simulate natural processes in order to predict or test for expected pattern. The sensitivity of stochastic spatial simulations to prescribed input landscapes can be evaluated by supplying them with a series of synthetic maps that obey particular statistical characteristics and then monitoring changes in modeled outputs. Statistically similar input landscapes with different spatial re-arrangements can be generated and supplied to spatial models as a hedge against pseudoreplication.
The quality of synthetic landscapes produced by the Fractal Realizer was tested using an online variant of the Turing Test. More than 1000 ecologists and mapping specialists were presented over the web with a series of 20 selections of paired maps, and asked to distinguish the real map from the synthetic realization from the Fractal Realizer. The resulting population of scores was not significantly different from a random binomial, proving that the experts were unable to discern the synthetic maps from the actual ones. Anyone can take the test at any time over the web at http://www.geobabble.org/realizer/turing.html
Since its publication in 2002, many thousands of people have taken the Turing Test of the Fractal Realizer. Several landscape ecology and GIS classes across the country have made the Turing Test of the Fractal Realizer a regular part of their scheduled laboratory exercises each year, and source code for the Fractal Realizer is available for download over the web. In April 2004, I was awarded the Outstanding Landscape Ecology Paper by the International Association of Landscape Ecology (IALE) for the publication in Conservation Ecology describing the Fractal Realizer (see Publication #46, 58 citations, http://www.ecologyandsociety.org/vol6/iss1/art2/)
In 2005, I developed MapCurves, a generalized algorithm for the quantitative, comparison of multiple categorical maps. MapCurves is a quantitative goodness-of-fit (GOF) method that unambiguously shows the degree of spatial concordance between two or more categorical maps. MapCurves graphically and quantitatively evaluates the degree of fit among any number of categorical maps and quantifies a GOF for each polygon, as well as for the entire map. The MapCurves method will even indicate a perfect fit between two ecoregion maps drawn by a “lumper” and a “splitter,” e.g., if all ecoregions in one map are comprised of unique sets of smaller ecoregions in the other map. It is not necessary to interpret (or even to know) legend descriptors for the categories in the maps to be compared, since the degree of fit in the spatial overlay alone forms the basis for the equivalency. MapCurves produces the best translation table between categories in each map as an output product, rather than starting with a guessed translation table as an input. Prior to MapCurves, meaningful quantitative comparison of two categorical maps was nearly impossible. One can compare two or more ecoregion maps using MapCurves, even if the maps contain radically different numbers of ecoregions. Two dozen well-known ecoregion and landcover maps were compared quantitatively using MapCurves (see Publication #59, 88 citations). One can also use MapCurves to “borrow” and apply the best, most appropriate labels from another map (of ecoregions or forest types, for example) to associate particular category names with each statistical quantitative ecoregion. MapCurves has been adopted by many others, and someone with whom I have no connection has written a downloadable R package called sabre: Spatial Association Between REgionalizations, which calculates MapCurves for other users.
Quantitative Ecoregions and LANDFIRE National Wildfire Biophysical
About 1997 I started experimenting with multivariate clustering as a way to statistically delineate homogeneous ecoregions, using a set of digital maps within a GIS as ecoregion characteristics. While recognizing the utility and popularity of ecoregions among ecologists and resource managers, I was dissatisfied with the reliance on subjective expert opinion used to produce them. I quickly realized that multivariate clustering represented a quantitative alternative that was transparent, objective, and repeatable.
Map cells are plotted in a high-dimensional data space using standardized values of each of their environmental characteristics as coordinates. Cells located close to each other must have similar mixtures of environmental characteristics, and perhaps should be classified in the same quantitative ecoregion. The number of ecoregions which result is under the user's control. Using closeness as a surrogate for similarity, an iterative classification procedure assigns every cell to the closest cluster centroid. After all map cells have been assigned, new cluster centroids are calculated to be the mean of each coordinate over all cells assigned membership to that cluster. Cluster centroids slowly move until the assignments converge and an equilibrium ecoregion classification is obtained. The algorithm was computationally demanding, mostly because of the large data volumes involved. These computational needs drove my research interest in pioneering construction of the Stone Soupercomputer from discarded personal computers (see Publication #40).
In 1999, I produced a national map of 1000 ecoregions created quantitatively by statistically clustering nine environmental variables, including physiographic, edaphic, and climatic variables at 1-km resolution. I also created a total soil Kjeldahl nitrogen map for the continental United States at 1-km resolution by combining non-agricultural data from the National Soils Characterization Database (NSCD) and STATSGO. I collaborated with others to link disparate tree physiology models to simulate tree growth across spatial scales (from leaf to stand to regions of stands), using forcing functions to drive models at larger scales (i.e., across the southeastern US, see Publications #32 and #44). Ironically, this Integrated Modeling Project was sponsored by the Southern Global Change Program, USDA Forest Service, which is now a part of EFETAC (see Publication #53, 272 citations.
Representativeness contours along a surface created by the distance from every cell to its clustered centroid can be used to sharpness or fuzziness of ecological borders, or ecotones, between ecoregions could be quantitatively characterized, even if it changes from side to side or along its length (see Publication #27, with 113 citations). The top three Principal Components of each ecoregion, when assigned to the three primary colors, create a unique set of statistical Similarity Colors for each ecoregion. These Similarity Colors show the degree of similarity or difference in the environmental conditions contained within each quantitative ecoregion. Coloring a quantitative ecoregion map with random colors emphasizes the edges between ecoregions, but the borders disappear entirely using Similarity Colors, and the map shows the dominant environmental gradients instead.
By submitting multiple maps of conditions occurring at different times to a single Multivariate Geographic Clustering process, a single set of common ecoregions are formed both across space and through time. Although the full set of ecoregions may not occur together within any single map, the same set of ecoregions occurs across the time series of all maps. Environmental conditions found within one ecoregion are the same, no matter wherever or whenever it occurs. Now quantitative statistical ecoregions can be followed through time, to see if they grow, shrink, join, appear or disappear. We call clustering through time and across space Multivariate Spatio-Temporal Clustering (MSTC), and it is particularly useful for tracking current ecoregions into one or more alternative predicted futures. Such through-time tracking is not possible when using human expertise-based ecoregions.
In 2003, I developed the first quantitative global ecoregion maps, sponsored by and in coordination with The Nature Conservancy (TNC). Single sets of quantitative ecoregions were statistically generated for both current global environmental conditions and future environments, as predicted for 2050 and 2100 by two global climate models under two possible future scenarios (http://www.geobabble.org/~hnw/global/ORNL-TNC/index.simcolors.html). In 2005, also with TNC, we developed the first quantitative ecoregion maps for Papua, New Guinea and China (see Publication #83, with 63 citations). These defensible and repeatable quantitative global ecoregions can be used to prioritize ecological conservation and restoration worldwide.
Using the same statistical quantitative ecoregion method, I was funded by the USDA Forest Service LANDFIRE project while still at ORNL to produce a set of National Wildfire Biophysical Settings Regions, based on 36 quantitative Wildfire-Relevant BioPhysical Characteristics, to map regions having similar burning conditions across the country for wildfire management (http://www.geobabble.org/~hnw/landfire/). Forrest Hoffman and I also designed and contracted a 136-node, 272-processor parallel supercomputer for the LANDFIRE, project, and most LANDFIRE products were produced using this parallel machine.
In 2013, Dr. Yasemin Erguner, a citizen of Turkey, was awarded a 1-year postdoctoral appointment, funded by the Turkish government, with the objective of producing a set of National Ecoregions for Turkey, under my direction, using Multivariate Geographic Clustering. We produced not only a series of the first quantitative Ecoregions for Turkey, but also mapped how those ecoregions would change under two alternative future conditions according to two leading Global Climate Models. The Turkish government was interested in the possible establishment of a National Turkish Ecological Sampling Network, similar to NEON in the United States, and the Turkish Ecoregions that we produced could form the basis for those nodes, just as our 20 Domains formed the basis for NEON nodes. This work resulted in Publication #109, .
In 2000, David Stoms and I argued that Normalized Difference Vegetation Index (NDVI), or “satellite greenness,” was a potentially important way to monitor vegetation health over wide areas (see Publication #29, 85 citations). Rather than clustering many different environmental characteristics, I began clustering repeated measurements of a single variable, NDVI, through time every 8 days for a full annual cycle. All locations having a similar annual profile shape of greenness timing would be clustered together into the same region. We named regions that shared the same NDVI phenology “phenoregions,” and produced a global map of phenoregions from AVHRR that might be used for monitoring climatic change (see Publication #56, 179 citations). We produced maps showing the 50 most-different national phenoregions, each having a differently shaped annual profile of greenness. These clustered phenoregion maps formed the basis for analyzing phenological trajectories of change in LanDAT (see Accomplishment # 12).
Parallel Multivariate Geographic Clustering has become one major focus of my research, and I have developed this quantitative approach into a rich and powerful quantitative statistical foundation that underlies many of my subsequent future achievements, including Scientific Accomplishments #4 (Network Analysis), #5 (Aquatic Invasives), #7 (ForeCASTS), #9 (Crop Mapping) , #10 (Fire Regimes), and #12 (LanDAT) listed here. Multivariate Geographic Clustering was involved in 45 of my 110 listed Publications, and continues to represent a thematic thread of continuity through my scientific career.
Analysis, Including AmeriFlux, FLUXNET and NEON Designs
When ecoregions are delineated using quantitative methods rather than expert judgment (see Scientific Accomplishment #3, above), the quantitative treatment provides a number of ecologically useful related concepts. Two of the most interesting of these are representativeness, which allows maps to be drawn which show the geographic location of all regions which are similar to a selected ecoregion (as used with ecological borders, above), and network site analysis, which shows how well a particular network of sites represents a larger area containing the network.
Quantitative ecoregions are of great practical use in the design and analysis of networks of installations or sample locations. Once input variables of appropriate relevance, scale and quality are chosen, the coverage and sampling intensity of any network of sites can be analyzed statistically with respect to those selected variables. Because the ecoregions are statistically derived, one can select a single ecoregion of particular interest, and then produce a sorted vector of the similarity of all other ecoregions to the selected one. Coding these pairwise similarity values as gray levels, a map can be drawn which cartographically shows the degree of similarity of all ecoregions in the map to the selected ecoregion of interest. Such maps, e.g., “Smoky Mountains-ness,” show the degree of innate multivariate similarity between a particular selected ecoregion and the rest of the map.
This similarity concept can also quantify how well an established network represents all of the conditions occurring within a map that contains it. A network consists of a geographic constellation of installations or facilities, or can simply represent locations where samples have been (or will be) taken. The quantitative similarity is now based on comparisons with multiple site locations within the established (or planned) network.
The best location for adding an additional new site or installation will be shown as the place that is the least well-represented by the current network of existing sites. Importance values for each site can be calculated, based on the marginal representation it adds to the network. Such importance values can be used to minimize the impact on representation if a site must be removed from the network. Finally, a network with a given number of sites can be designed which is theoretically optimum, having the highest possible representation on the map.
Until now, sites in even large-budget networks have been established in opportunistic, political, or logistically-driven ways, resulting in undirected, organic growth. Network analysis is simple, quantitative and defensible, and provides the first objective guidance for network design and evaluation (http://www.geobabble.org/~hnw/networks/).
I initially used this approach to determine the degree to which the existing network of carbon eddy flux towers within the AmeriFlux network are representative of flux environments across the conterminous United States (http://www.geobabble.org/~hnw/networks2/). This network representativeness information was used to determine how many additional AmeriFlux towers will be required, and where additional towers should be placed. In addition, the importance and uniqueness of each existing tower to the Ameriflux network were calculated. This quantitative ecoregion-based approach to stratifying carbon flux may be the fastest way to fulfill the North American Carbon Program (NACP) and AmeriFlux goal of seasonally mapping sources and sinks of carbon within the North American continent (and see FLUXNET2015 work, below). Sponsored initially by the Office of Biological and Ecological Research (OBER), DOE, I received additional funding from AmeriFlux to continue this AmeriFlux network analysis, resulting in Publications #48 (80 Citations) and #50.
After 8 years of additional tower site additions and losses within AmeriFlux, Beverly (Bev) Law, the Director of AmeriFlux, asked me early in 2011 if I would repeat this analysis for the current configuration of the AmeriFlux network. Now working in the Forest Service, I repeated the AmeriFlux network analysis, and updated representativeness results were presented at the 2011 annual AmeriFlux meeting.
In 2016 we used network analysis to calculate global representativeness of the FLUXNET network of flux towers, showing regions which were poorly represented by the current geographic constellation of operating FLUXNET eddy covariance towers. The FLUXNET2015 dataset (released in late 2016) contains global FLUXNET measurements from member eddy-covariance flux towers located all over the earth. We used our Generic Imputer (see Accomplishment #7) to produce monthly global maps of ecosystem Gross Primary Productivity for 20 years, producing planetwide monthly maps of GPP from upscaled flux tower measurements. Paper was submitted to Earth System Science Data journal and received favorable peer review (see the full story under “Publication” #108).
In 2018, Alisa Coffin contacted me, asking if we could use our network analysis methods to calculate the representativeness of their Long-Term Agroecosystem Research Network (LTAR). We have calculated national maps of LTAR Network Representativeness, and LTAR Network Constituency, based on ecological growing conditions, but LTAR also wishes to include socio-economic variables and crop productivity data, in order to gauge representativeness with respect to Cropland, Grazingland, and Integrated Systems. The representativeness analyses may be used as a way to upscale measurements, and as an argument for funding additional LTAR sites in poorly represented locations. Our initial representativeness maps received an Impact Award at the recent LTAR national meeting in June.
Because of the development of these network design and representativeness capabilities, I became involved with the early design of National Ecological Observatory Network (NEON). NEON is the first ecological measurement system designed to answer regional- to national-scale scientific questions. A system of identical nodes was envisioned, each representing the ecological environments within the United States. All nodes are focused in unison on a few transformational ecological questions of national relevance. To better sample the diverse ecological environments of the United States, those environments were first divided into a set of more homogeneous "strata." NEON nodes could then be located within each stratum, helping to ensure that their measurements can be scaled up to represent the entire range of environments within the United States. Multivariate clustering based on national maps of 9 ecologically relevant climatic "state" variables was used to repeatably define 25 national climatic zones. These 25 climate zones were combined with dynamic air mass seasonality data to create 20 NEON domains, each having relatively homogeneous climate.
I was invited to become a member of the 15-person NEON National Network Design Committee (NNDC), which met two dozen times over a period of 4 years. The NNDC was responsible for drafting the Integrated Science and Education Plan, the Networking and Informatics Baseline Design, and the Project Execution Plan (PEP) for NEON. I was also involved in the NEON Conceptual Design Review (CDR). These reports and review results were given to the National Science Foundation and the U.S. Congress, which funded NEON construction.
Using quantitative ecoregions and optimal network design, I suggested the national regionalization on which the 20 official NEON domains are now based (http://www.geobabble.org/~hnw/neon/neonindex/). This summary article on NEON in Science mentions me by name, and the NEON website describes my contribution to the development of the official NEON domains. These efforts resulted in Publications #65 (39 citations) and #69 (220 citations).
The 20 NEON domains are fundamental, underlying everything else in the NEON network. NEON is the largest and most expensive environmental project that the National Science Foundation has ever undertaken. The NEON network has now been completed at a cost of more than half a million dollars, and NEON will have a lifespan of 30+ years. These are among my most significant and lasting scientific contributions, and likely represent the summit achievements of my scientific career. I currently serve as a member on the NEON Spatial Sampling Technical Working Group (TWG).
We were funded by the Office of Biological and Environmental Research (OBER) within the Department of Energy’s Office of Science to use our network analysis to analyze DOE’s two new Next Generation Ecosystem Experiments (NGEE)-Arctic Alaskan Climate Change ecosystem warming sites, to study the placement of the sites within Alaska, and to estimate the representativeness of their measurements. NGEE-Arctic is a major 12-year, $20M DOE research effort. Publication #92 that resulted was selected as the Outstanding Paper in Landscape Ecology by US-IALE in 2014.
Species Predictions for the Great Lakes and Sudden Oak Death
In 2009, my student Matt Fitzpatrick and I developed a model “transplantation” method to first develop a niche model the invasive fire ant, Solenopsis invicta, within its native habitat and apply the model to the invaded lands, and then to develop a second niche model within the invaded habitat and use that to project home range in its native habitat. Although the invader may not yet have colonized the full extent of invaded lands, the differences quantify the degree of release from native predators that the invader has enjoyed in the new area. We were also among the first to name and discuss the challenge of “non-analog” future climatic conditions (see Publication #79, 271 citations).
About the same time, hired as a consultant to the Environmental Protection Agency (EPA), I predicted the exotic aquatic organisms most likely to invade the Great Lakes from the Ponto-Caspian Sea region. I identified the most likely aquatic invaders across all taxa, and predicted the geographic extent of the potentially susceptible areas for each species within the Great Lakes.
Aquatic invasive species are transported with normal ship traffic, often carried in ballast water. This study predicted susceptibility by quantifying the degree of multivariate similarity of aquatic environments worldwide to selected locations within the Great Lakes, USA. The approach assumes that, sooner or later, transport of invasive aquatic organisms will occur to and from all points on the globe. Following such human-mediated accidental transplantations, it is the degree of similarity of the new aquatic environment to the original environment that determines whether the invader will successfully establish a population in the new location.
We produced multiple sets of aquatic ecoregions, based on six characteristics of the surface aquatic environment. Using Multivariate Geographic Clustering (MGC, see Accomplishment #3) on a parallel supercomputer, we grouped over 50M 4 km map cells into groups or clusters having similar combinations of the six environmental conditions. When placed back into geographic map space, these groups form geographic regions across all global aquatic habitats which share similar environmental conditions (http://www.geobabble.org/~hnw/global/aquaticinvaders/, aquaticinvaders2).
As with maps showing “Smoky Mountains-ness” (see Accomplishment #3), , a world map can be drawn in which the degree of multivariate similarity between the aquatic environment in the selected location and the aquatic environment in every other location is shown as a shade of gray. By quantifying the similarity between aquatic environments, such maps show both the locations from which aquatic invasive organisms that are likely to survive here might come, and locations to which invasive aquatic forms from this location might go and establish a viable population (aquaticinvaders3 and aquaticinvaders4). Using this similarity-based approach to map global oceans and lakes into aquatic ecoregions, it is not necessary to select particular donor and recipient locations, nor to do the analysis on a tedious species-by-species basis (see Publication #76). A similarity approach is also being used to develop global Invasibility Zones for terrestrial invasive species (see Accomplishment #7, bottom).
The same quantitative ecoregionalization-based process has proven useful for mapping the risk that Sudden Oak Death (SOD), Phytophthora ramorum, will spread to other parts of the U.S. The susceptibility of forests beyond the west coast of the United States to SOD is unknown, but is the subject of speculation, since the spread of the SOD epidemic could represent a serious threat to eastern forests. I created custom statistical SOD-relevant ecoregions using national maps of conditions likely to be limiting for P. ramorum, including humidity, leaf-wetness, and cool temperatures. My analysis of the quantitative multivariate similarity of each of these 1500 homogeneous SOD-regions with conditions in known SOD outbreak areas produced a continuous national estimate of risk or susceptibility to SOD.
Practical Map-Analysis Tool for Corridor Detection
I led the development of a landscape map analyzer tool which will identify and map corridors and barriers to plant and animal movement across any map. Corridors are the "roadways" most commonly used by plants and animals as they move or disperse across a mapped landscape. The tool is based on the idea of island biogeography, and considers the map as isolated patches of high-quality habitat embedded in a matrix "sea" of all other patches of lower quality.
Corridor connectance, whether we wish to preserve it for a threatened species or impede it for an invasive exotic, is a critical concept in biodiversity management. Despite this importance, the idea of corridors remains largely conceptual. Few analytical management tools exist which can examine a real-world map, quantify connectance, and identify potential corridors.
The tool we developed, called “Pathway Analysis Through Habitat,” or PATH, uses simulated virtual plant and animal "walkers" which are imbued with movement characteristics and preferences of particular animal or plant species, and allows large numbers of these imaginary digital walkers to travel over the map. Virtual walkers representing individuals "try" to successfully disperse from one "island" patch of favorable habitat to another "island" in the archipelago, and, in so doing, define and map the best potential movement dispersal corridors. The spatial arrangement and amount of patches of habitat, roads, urban areas and other real-world landscape features will affect the movements and successful dispersal of walkers, and therefore the routes of potential corridors across the map.
PATH provides realistic guidance for conservation and management decisions. PATH patch importance values can be used to direct and prioritize planning for conservation and remediation. A land manager can easily see which habitat patches are the most important targets for conservation or strenuous remediation (for a threatened species) or for elimination (in the case of an invasive species). In the case of an invasive species, patches important as connecting corridors, once identified, would be the first places that a manager would want to make inhospitable for the invader. Construction of unsuitable or barrier patches, or elimination of particular favorable "bridge" patches may be suggested which will discourage movement of invasive species along existing corridors. For threatened species, patches with high importance should be vigorously protected or preferentially remediated; while patches with low importance are more available for alternative use or development.
The PATH tool was developed using funding from the Southern Appalachian Information Node (SAIN) of the National Biological Information Infrastructure (NBII) of the US Geological Survey, the Army Corps of Engineers ERDC-CERL, and the National Petroleum Technology Office, Department of Energy. The prototype was initially run on small artificial test maps to evaluate its behavior for simple artificial landscapes designed to produce expected intuitive results , and then simple actual landscapes (Publication #60, with 94 citations). As a parallel application running on a supercomputer, the PATH tool is computationally powerful enough to analyze movements of large megafauna across extensive, highly-fragmented multi-state real-world landscapes. A CERL Technical Report, for example, describes how the PATH tool was used to analyze Red-cockaded woodpecker movement across the Southeastern United States (Publication #62). A manuscript showing potential gopher tortoise movement corridors within and around Fort Benning, GA using the PATH tool is being prepared for publication (http://www.geobabble.org/~hnw/walkers/gophertortoise). Because of the level of interest shown by the US Armed Forces in analyzing connectance, I worked with ERDC-CERL to translate the PATH tool into the NetLogo language, so that PATH no longer requires the use of a parallel supercomputer, thus making it more accessible to resource managers (Publication #89).
Tree Species Range Shifts Under Two Alternative Climate Change
In 2005 we produced a set of global ecoregions through time with The Nature Conservancy (TNC) to use as a basis for climate change conservation triage, based on climatic shifts projected from the Hadley model under two alternative scenarios for the United States in 2100 (Publication #57, with 117 citations). Environmental domains found across half of the study area today disappeared under the higher emissions scenario. Areas at lowest risk which represented potential refugia, and areas at greatest risk allowed TNC to prioritize particular areas for conservation.
Climate change poses a severe threat to the viability of several forest tree species, which may be forced either to adapt to new conditions or to shift their ranges to more favorable environments. Species already having limited geographic ranges may be at highest risk. Along with Kevin Potter, I used spatial models of future environmental conditions to predict future suitable geographic range shifts for several hundred tree species under different climate change models and emissions scenarios. We also determined where each species, within its current range, is most susceptible to local extirpation as a result of climate change.
We used the predictions from two Global Climate Models, with two climate scenarios each, for two future dates, plus present conditions (nine copies of the earth at 4 km2 resolution), in a single Multivariate Spatio-Temporal Clustering (MSTC, see Accomplishment #3) on the supercomputers at ORNL to statistically create 30 thousand global quantitative “Suitability” Ecoregions through time, formed on the basis of 17 ecological variables describing temperature, precipitation, soil and topographic characteristics. MSTC identifies the same 30 thousand ecoregions across all nine Earths, so that these ecoregions can be tracked into each alternative future. We used Forest Inventory Analysis (FIA) plots (United States only) and Global Biodiversity Information Facility (GBIF) Data (Worldwide) as Occurrence points to find the subset of the 30,000 ecoregions within which this tree species can survive. This subset of ecoregions comprises the present-day suitable home range for this species. If we use MSTC to track this subset of suitable ecoregions into the future, does this tree’s future range move, shrink, grow, overlap the present range, or vanish?
Ecoregions containing a species occurrence point are colored red, delineating its current suitable range. Ecoregions without an occurrence point are colored in shades of gray that shows their degree of similarity to the most-similar occupied ecoregion, based on the quantitative multivariate similarity across all 17 environmental conditions. There is no species-specific "tuning" at all, enabling rapid climate change assessments to be done quickly for many tree species. Each tree range was predicted with and without elevation.
Range predictions can be evaluated by how well the predicted current range matches the known current range for that tree species. Elbert Little, Chief Dendrologist, USDA Forest Service, 1907-2004, published the Atlas of United States Trees, containing hand-drawn maps of geographic ranges for most tree species. Little's maps are still the best maps we have for tree ranges. When comparing a predicted suitable (or fundamental) current range with Little's actual (or realized) tree range maps (which are approximations themselves), Little's range should be slightly smaller, since the realized range is geographically squeezed by competitors, predators, and parasitoids.
In the Forecasts of Climate-Associated Shifts in Tree Species (ForeCASTS) project, range shifts for 325 tree species were predicted globally following future climate changes forecast by the Parallel Climate Model (PCM) and the Hadley Climate Model under IPCC scenarios A2FI and B2 for the years 2050 and 2100 ( ). All but a handful of tree species’ predicted present ranges closely match Little’s maps. Most exceptions, like chestnut, have reasonable explanations for differences. Because there has been as much interest in the Present-Day Range predictions as in the predicted future ranges, the predicted current “Hargrove” maps are downloadable as GIS files for each of the 325 tree species.
Minimum Required Movement (MRM) Distance determines how far a species would have to move in order to arrive at the nearest location with the same combination of conditions they had prior to a climatic change. Global maps showing MRM distance to return to the closest geographic locations offering suitable conditions in the future directly show the likelihood of local extirpation following climate change. Locations that are the nearest "lifeboats" for large surrounding areas may represent management and conservation targets.
Version 4 of the ForeCASTS Species Atlas, made available to managers in 2012, contains predicted future host range maps for more than 325 tree species, covering essentially every woody species whose home range currently extends into the conterminous United States. Resource managers, land-use planners and conservation organizations can view ForeCASTS future host range maps for any U.S. tree species at https://www.geobabble.org/ForeCASTS/atlas.html. Unlike existing tree range shift prediction atlases, which are limited to the eastern or western United States, ForeCASTS maps are global in extent. With maps for 325 tree species, ForeCASTS already covers many more tree types than earlier tree-shift climate change efforts. Results for several tree species were used in planning for several NFs (Francis Marion and Sumter NFs) and several states (NC, Linda Pearsall, NCDENR, and WA and OR, Carol Aubry, USDA FS, Olympic NF). A poster showing ForeCASTS results received the “Most Exciting Science” Award at the Forest Service Forest Health Monitoring (FHM) Work Group meeting in April 2010. ForeCASTS species range shift results were used in “A Mid-Atlantic Forest Ecosystem Vulnerability Assessment and Synthesis” (General Technical Report NRS-181, October 2018). In addition to the ForeCASTS website, future climatic risk results were reported in Publications #80, #90, #94, #100, and #104 (explained in audio here).
Once independently predicted, the 325 future tree ranges can be stacked and subjected to higher-order analyses. For example, the top and bottom 20 generalist and specialist tree species can be quantitatively ordered by niche breadth, using the number of the 30 thousand Suitability Ecoregions within which they occurred. Similarly, national maps of current Tree Species Richness show that the present-day center for Tree Species Richness is in central Alabama, but the center of future Tree Species Richness moves to central Georgia. Tree Species Endemism, which can be quantitatively calculated using not just “rare” species but all modeled trees, is often used as a surrogate for habitat conservation importance. The current hotspot for Tree Species Endemism moves from the Blacklands of Alabama to central Georgia by 2050. Such higher-order results are not possible without first making the individual species-by-species forecasts and summing the results.
Jitendra Kumar and I developed a “Generic Imputer” to estimate continuous gridded maps of species productivity. The tree ranges in ForeCASTS are binary – either suitable or unsuitable – with no estimates of growth or productivity. A 300-year old tree might survive, but be small, with slow growth. We have sparse measurements of productivity at the FIA plot locations, but would like to "spread the measurements out" into a continuous gridded map throughout the range.
To impute a productivity surface across the entire range for present and future conditions, we "associate" some sparse FIA measurements of productivity with each clustered Region, but these measurements are NOT used in the clustering. Imputation is done in data clustering space, NOT in geographic map space. Regions are large, and there is variability in fertility across micro-sites within the same region. We use the 90th percentile of all growth measurements within a Region. If there are no measurements in a Region, the Generic Imputer uses multiple values from the next-most-similar (closest in data space) regions that have measurements.
We produced an even finer 20 thousand Productivity Ecoregions within the CONUS for the imputation of tree species-specific continuous national Productivity Surfaces. Importance Value is a productivity measure that integrates the frequency and density of individuals with basal area growth. The Generic Imputer uses sparse Importance Value measurements made at FIA plots as sparse data input for imputation of continuous national Productivity surfaces for a subset of tree species with average-sized ranges. We have also used the Generic Imputer on the new FLUXNET2015 dataset to impute continuous gridded global monthly maps of ecosystem Gross Primary Productivity from upscaled flux tower measurements for 20 years, from 1991 to 2014 (see “Publication” #108).
With Kevin Potter, I employed the ForeCASTS methodology to predict quantitative Seed Transfer Zones, within which seeds can be transferred from local sources, planted, and expected to have suitable growth. The importance of seed sources has been long understood for success in reclamation and recovery efforts, but Seed Transfer Zones have been mapped only qualitatively and coarsely, since no quantitative mapping methods existed. ForeCASTS allows quantitative mapping of species-specific Seed Transfer Zones for two distinct types of uses, Forward and Reverse. Forward:“If I have seeds from a given location, where can I plant them to best ensure the trees will be well-adapted in the future?’’ Reverse:‘‘If I want to plant trees in a given location and to ensure that those trees will be well-adapted in the future, where do I go now to collect the seeds?’’ Even the separation of these two “Seeds from here now, plant where for later?” versus “Trees for here later, seeds from where now?” questions as distinct efforts stems from this work, and the answers are often not reciprocal. In 2012 we published Determining Suitable Locations for Seed Transfer under Climate Change: A Global Quantitative Method. (Publication #90, 49 citations) describing these quantitative Seed Transfer Zone methods, showing Forward and Backward examples for Pinus palustris and for Cornus florida.
I am extending the ForeCASTS methods to develop global “Invasibility Zones” to rank relative dangers from Invasive Species. I had already used similar methods to map Phytophthora ramorum invasive susceptibility nationally and explore Eastern sensitivity to Sudden Oak Death (see Accomplishment #5, bottom), and have used multivariate clustering to form Global Aquatic Ecoregions to gauge the susceptibility of ports in the Great Lakes to aquatic invasive species (see Publication #76 and Accomplishment #5). Most traditional approaches to invasive species, like the one my student Matt Fitzpatrick and I used to predict areas susceptible to invasive fire ants (Publication #79), have been species-specific, involving complex niche modeling methods repeated for each possible species threat before the full risk could be estimated. Often a species not even originally foreseen as a risk becomes the worst, most successful invader.
General analysis of the overall similarity of environments between multiple locations could replace this slow, stepwise species-by-species approach. Assuming eventual cosmopolitan transport of all species, propagules should be more likely to successfully establish if their new environments are very similar to the home environments from which they came, in the same way that the Seed Zones method worked above. I am establishing 8 to 10 global Invasibility Zones which quantitatively map similar environments. Two or more member locations within the same Invasibility Zone must fastidiously guard against exchanging Invasive Species propagules with each other, because of the great similarity of their environments. Two locations that are members of different Invasibility Zones need not be so careful or concerned, according to a decreasing sorted quantitative similarity list of Zones that are produced.
Invasibility Zones are general and species-free, and show not only the concern for receiving successful propagules, but also the reciprocal risk of sending successful propagules to other global locations. Invasibility Zones maps and tables could be used as thumbnail guides to help, e.g., overwhelmed APHIS inspectors to rationally divide available inspection efforts among multiple simultaneously arriving container ships.
Global environmental Invasibility Zones can easily be created with Clustering using the ForeCASTS data, but they must be calibrated using some external standard for how similar invaded and native home environments need to be to permit establishment of invasive species. The Global Naturalized Alien Flora (GloNAF) database is the first database on alien vascular plant species distributions worldwide. First published in Nature in 2015, GloNAF includes 13,939 taxa and covers 1,029 regions, with information on whether the invasive taxon has become naturalized and self-sustaining. EFETAC signed an MOU with GloNAF in order to use the data set for the development and calibration of global Invasibility Zones.
in the Sky” Early Warning System Monitors Forest Disturbances
EFETAC was created in 2005 under a Congressional Mandate to develop a national-scale Early Warning System for forest disturbances. Along with our collaborators, I conceived and established the ForWarn National Early Warning System ( , Publication #78, 68 Citations, describing the custom forest damage algorithm), which produces a suite of maps showing forest disturbance across the United States at 231m resolution every 8 days (view introductory video [large download!]). ForWarn (https://forwarn.forestthreats.org) is an on-line, near real-time satellite-based forest monitoring and assessment tool for detecting and tracking potential disturbances in forests across the North American continent. ForWarn provides new forest change maps every 8 days for most of North America, even throughout the winter. Since January 2010, the ForWarn system has been used to detect environmental threats to forests caused by insects, diseases, wildfires, extreme weather, and other natural and man-made events. “Departures” that can be detected include not only classical forest disturbances like insects, disease, and wildfire, but also the effects of inter-annual weather deviations, including extremes of temperature and precipitation, making vegetation responses to heat, cold, flooding and drought easily viewable. The frequent updates produced by ForWarn allow forest managers to take more responsive, effective forest management actions, and to track recovery in forests following disturbances. No such national-scale system based on remote sensing has been developed specifically for forest disturbances before. ForWarn was the result of an ongoing, substantive cooperation among four different government agencies: USDA, NASA, USGS, and DOE, and the Federal Laboratory Consortium (FLC) honored the ForWarn team with its 2013 Interagency Partnership Award, one of the highest honors from the FLC. ForWarn is currently finishing its eighth year of operation.
ForWarn detects most types of forest disturbances, including insects, disease, wildfires, frost and ice damage, tornadoes, hurricanes, blowdowns, harvest, urbanization, and landslides. It also detects drought, flood, and temperature effects, and shows early and delayed seasonal vegetation development. Cells in the map are about 5 ha, or 13 acres each. ForWarn works by comparing current greenness with the “normal” greenness that would be expected for healthy, undisturbed vegetation growing at this location during this time. Locations that are currently less green than expected are identified as potentially disturbed. A set of five disturbance products use differing lengths of historical baseline periods to calculate the expected normal greenness, highlighting how recently the forest disturbance has occurred. An "Early Detect" product returns the most recent cloud-free NDVI observation, providing forest managers with the earliest possible initial indications of new forest disturbances.
ForWarn products can be viewed by anyone using the online Forest Change Assessment Viewer (http://forwarn.forestthreats.org/fcav2), which runs on any computer with a web browser; no special programs are downloaded to the machine, and no user IDs or passwords are required. The interface is intuitive and familiar, similar to Google Maps. The Assessment Viewer contains all current and historical ForWarn maps, along with co-registered maps of insect and disease outbreaks, wildfire perimeters, and much additional disturbance-relevant information. Using a Share-this-map feature, users can paste and send a URL that, when clicked on by others, launches the Assessment Viewer showing them exactly the same ForWarn disturbance map, facilitating consultation with the ForWarn Team. Members of the Team use the same Assessment Viewer tools that are available to ForWarn users, and users see the latest ForWarn maps at the same time as Team members do.
During the growing season, the ForWarn Team notifies federal, state, and private forest health professionals when alerts are warranted. A warning email (containing a “Share this Map” URL to the ForWarn Assessment Viewer) is issued by the ForWarn Team to one or more local and regional resource managers, allowing them to identify and track the forest disturbance. We selectively alert entomologists about insect disturbances, and plant pathologists are alerted about forest diseases, while forest owners and Regional FHM Coordinators receive all ForWarn disturbance notifications. In many cases (e.g., Atchafalaya, LA in 2010 and 2012, forest tent caterpillars and bald-cypress leafrollers, and Allegheny NF, PA in 2011, fall webworms), ForWarn has alerted local resource managers to otherwise unknown insect defoliation activity. In the Atchafalaya 2010 and 2012 cases, an extra, unplanned IDS flight was made which verified the defoliation. The Allegheny NF defoliation was verified by ground observations. ForWarn mapped many tornadoes, wildfires, extreme drought, and insect defoliations during the 2011 growing season. (see Publications #93 and #98, also the article in Space News, April 2012, and the Capital Ideas - Live! interview). Over 300 ForWarn alerts have now been issued nationally, for many causative agents.
The ForWarn system has become extremely popular, enjoying a groundswell of support from federal, state, county and private foresters, and earning several prestigious awards. Prior to ForWarn, forest owners and resource managers relied solely on the USDA FS Insect and Disease Survey (IDS) Program to provide annual regional geospatial data on forest conditions and trends. IDS utilizes aerial “sketchmappers,” who identify and map apparent forest disturbances from light aircraft using hand-held Geographic Information Systems. IDS data are collected regionally, and then "rolled up" into a single national coverage released the following growing season. IDS work is conducted by highly trained specialists, but is subjective, time-consuming, hazardous, incomplete, and costly (averaging $14M per year). For example, IDS only mapped 70% of the CONUS forests in 2012; this consisted of a single overflight for the entire calendar year; some forested areas receive no disturbance monitoring at all. With ForWarn, forest managers are afforded the luxury of postponing difficult forest management decisions by simply waiting 8 days to see the next set of national ForWarn disturbance maps.
The ongoing ForWarn Detect/Warn cycle builds a network of continually growing partnerships between the ForWarn Team and working forest resource professionals everywhere. Once forest managers have received and verified a ForWarn alert for a disturbance detected in their own forests, they usually become committed ForWarn users themselves, carefully watching all future ForWarn products faithfully. For example, an invited talk was given about ForWarn at the 2013 Intertribal Timber Council Meeting, and then a Memorandum of Understanding (MOU) was signed between the Menominee Nation and EFETAC. In this way, ForWarn continues to establish lasting two-way partnerships with ever-increasing numbers of forest managers across the United States, looking “over their shoulders” as they use ForWarn themselves to find, identify, and verify forest disturbances within their own forests. In 2012, the ForWarn Team received the Southern Research Station Director’s Award for Science Delivery, and in December 2013, the ForWarn Team received the Chief’s Award from Thomas L. Tidwell, for “helping to preserve and enhance the nations forests and grasslands” (award cover letter).
From Jan 2010 thru April 2017, our ForWarn colleagues at NASA Stennis Space Center, who actually calculated the products, were never late with a product delivery date. However, in 2016, Stennis unexpectedly changed its scientific mission away from Applied Science, and most of the ForWarn-related personnel moved to Leidos, Inc. This employment shift necessitated that the Research Ecologist establish a new, sole-source contract, which resulted in a gap in ForWarn products for about a year. About this same time, the USGS eMODIS data used as the source for ForWarn changed, no longer including information from both MODIS sensors. No ForWarn maps were produced for the 2017 growing season, but production resumed in April 2018, and has been uninterrupted since.
The Research Ecologist took advantage of this one-year production hiatus to re-engineer and improve many aspects of the ForWarn system. The new system was called ForWarn II, indicating similarity to users, while earmarking major improvements in source data, production methods, products and extent. He located and implemented a new data source, the NASA Goddard Spaceflight Center GIMMS/GLAM (Global Agricultural Monitoring) System as a new alternative input data feed for the new ForWarn II system. GIMMS/GLAM uses Collection 6 for both MODIS sensors, is partially funded by USDA, and has a geographic coverage that extends globally, beyond the conterminous United States. The Research Ecologist personally wrote and tested more than 5000 lines of code using the Geographic Data Abstraction Library (GDAL), ingesting GIMMS/GLAM data and devising a new method of production which is independent of any GIS system, and requires no proprietary software packages having annual re-licensing costs.
The Research Ecologist’s ForWarn II production code utilizes virtual Cloud Computing, purchasing computing cycles as a service, obviating the need for EFETAC to purchase, maintain and update actual physical computer hardware. The production code automatically downloads all necessary MODIS data from GIMMS/GLAM every 8 days, but can also use provided precursor files to shorten computation run times. The new ForWarn II production codes allow closer Forest Service control while greatly decreasing production costs, enabling a longer and more likely continued future lifespan for ForWarn II.
Forest insects and diseases know no political boundaries. ForWarn started in 2010 with a lower-48 state CONUS spatial extent. Using the new GIMMS/GLAM input data, ForWarn II resumed in early 2018 with extended spatial coverage, including boreal Canada, Mexico, and the Caribbean. A third new increase in extent is now underway, adding coverage for all of Central America, Hawaii, and Alaska. Increased snow and cloud cover necessitated development and use of different processing methods in Alaska than are used elsewhere. This final enlargement will give ForWarn II a truly continental North American perspective on forest disturbances.
ForWarn II makes some changes in the standard 1-, 3-, 5-, 10-, and all-year ForWarn products, and adds three new products, Seasonal Progress, Disturbance Duration, and Disturbance Rank. The Research Ecologist also used his new production codes on supercomputers at ORNL to back-calculate the complete historical archive of all ForWarn products backwards every 8 days to January 2003, the beginning of the MODIS period. These historical products are available in the viewer, so that a forest manager can review how any pre-2010 historical disturbance in their forests would have appeared in ForWarn.
During 2019, Leidos, Inc. maintained parallel production of ForWarn “Legacy” products, allowing extended comparisons with the new ForWarn II production line, but this duplication will end next season. Thus, the new in-house production system for ForWarn II has already saved EFETAC $30K of FY2019 funds, and will ultimately permit EFETAC an annual production cost savings of nearly $215K, representing the major portion of one existing ForWarn subcontract.
The 35-day 2018-19 government shutdown precluded an official release of ForWarn II in April 2019. The new cloud-based ForWarn II production codes continued to automatically produce ForWarn II products unattended throughout the shutdown period, although there were no Rapid National Assessments, and no disturbance alerts were issued. Official release and public rollout of continental-extent ForWarn II is now planned to occur before the end of the 2019 growing season.
Agricultural Crop Mapping to Permit Agricultural Monitoring and
Detection of Crop Disturbances
ForWarn tracks disturbance in all vegetation, not just forests, including potential disturbances in rangeland vegetation and agricultural crops. This all-vegetation feature of ForWarn may widen the potential user audience to include farmers as well as forest owners and range livestock managers. Unlike forests that (usually) remain growing in the same places from year to year, farmers often plant different crops in the same field, using an unpredictable rotation system. ForWarn already monitors agricultural vegetation, but it assumes that, like forests, the same commodity is planted this year as in prior years. If the crop this year has been changed, normal greenness that is used for comparison will be inappropriate, and the relative crop health status shown by ForWarn will be incorrect. However, if ForWarn could be provided a map of crop types planted in this current growing season, it could be used to monitor crop health nationally every 8 days along with forests and rangelands. USDA produces a national Crop Data Layer (CDL) annually showing the location of all crops, but the CDL is not released until the following growing season, too late for current-season use by ForWarn.
In 2008, I worked with Carol Williams at Iowa State to produce and publish crop ecoregions of Iowa (Agro-ecoregionalization of Iowa using Multivariate Geographical Clustering, Publication #68, 60 citations). I am currently helping direct a Ph.D. student from Northeastern University, Venkata Shashank Konduri, whose work on national within-season crop identification was partially sponsored using EFETAC ForWarn II funds. Using only my Clustering and my Mapcurves methods (see Achievements #3 and #2) on 8-day MODIS NDVI, Shashank has now achieved national crop identification accuracies of more than 65% at 30m resolution for each of the 8 top commodities by area planted. When summed to counties, these accuracies approach 90% for the commonest crops. Crops have achieved 90% of spatial mapping accuracy by mid-July for corn and winter wheat, within the same growing season. A manuscript for Remote Sensing of Environment is pending.
If it can be made as useful for agriculture as it has been useful for forestry, ForWarn II may find alternative agriculture-based funding sources for continued operation. We visited Rick Mueller, who produces the CDL annually at the USDA National Agricultural Statistics Service (NASS) and USDA Risk Management Agency (RMA) to see if ForWarn results can be leveraged elsewhere within our own agency.
Determined Global Fire Regimes
My research with wildfire started with my 3 seasons of field research in Yellowstone and my EMBYR wildfire model (Publication #36, 306 citations, ). Bob Gardner and I used EMBYR to simulate the wildfire regime over the next millennium for the Greater Yellowstone Ecosystem under three alternative, synthetic, fractally-generated climate scenarios, with 10 replications of 1000 years into the future for each of the three climate scenarios (see Publication #17, with 131 citations). For LANDFIRE, I used Multivariate Geographic Clustering (see Accomplishment #3) of 36 Wildfire-Relevant BioPhysical Characteristics to produce a National Map of Wildfire Biophysical Settings, regions having similar burning conditions across the country for wildfire management (http://www.geobabble.org/~hnw/landfire/). Ostensibly, wildfires would have similar burning conditions occurring anywhere within any one of the 3000 Biophysical Settings regions.
The FSIM National Wildfire Probability Map, produced by Mark Finney at the Missoula Fire Laboratory is widely used as an index of local wildfire risk. Yet, as the product of a complex simulation model requiring thousands of hours of computer time, the FSIM Map is difficult to judge or evaluate. The FSIM Map represents such a huge effort that it is a challenge to produce other, additional independent efforts with which to compare or corroborate it. In 2015, we used two of our existing products to perform a more observation-based, hypothesis-free empirical and independent comparison check of the National FSIM Map.
We compared where wildfires have historically occurred over the last 30 years (MTBS wildfire perimeters) with two categorical maps that we statistically produced using direct remote sensing observations - one map representing Fuel/Vegetation Types, and one map of Wildfire Burning Conditions/Biophysical Settings. We used our clustered National MODIS Phenoregions map, after stealing fuel type labels from the LANDFIRE Fuels Map using MapCurves, as our National Fuel/Vegetation Types map. We overlaid historical MTBS Wildfire Perimeters to produce a ranked, categorical map for each, colored by rankings, and compared the results with Finney's FSIM Probabilities Map. Finney's FSIM Map results were largely supported by consensus with these independent probability maps. There are a few consistent regional differences, and more FSIM commission differences than omissions. Comparison results were presented at the American Fire Ecology (AFE) meeting in San Antonio, TX in 2015.
These preliminary observational and modeling studies of wildfire environments and fuel settings led me logically to the quantitative and empirical consideration of global Fire Regimes. Fire Regimes are geographic regions within which wildfire occurrences have similar repeating patterns of burn intensities, return intervals, and seasonality. By delineating regions that share common wildfire characteristics, Fire Regimes can show additional locations where particularly successful wildfire management or response strategies can also be used, or where methods tried elsewhere unsuccessfully are also unlikely to work. But all existing Fire Regime maps have been drawn subjectively, using only expert opinion and existing conceptions.
We used thermal “hotspot” data collected globally by the two MODIS sensors during four overpasses per day/night throughout their 17-year orbital history in the Multivariate Geographic Clustering process (see Accomplishment #3) in order to statistically produce a quantitative discrimination of different Fire Regimes globally, including identification of similar regimes across hemispheres. We included both human-caused fires and wildfires, classifying both types of Fire Regimes empirically.
To appropriately address opposing seasonal juxtaposition across northern and southern hemispheres, I developed a special transformation of fire dates, based on latitude and temporal proximity to solstices and equinoxes, which allows statistical discrimination of, say, “summer” fires, regardless of the calendar month or hemisphere in which they occurred. This new date transform permits recognition of similar fire seasonality in both northern and southern hemispheres. Representation of day-of-year as sine/cosine pairs allows the clustering algorithm to recognize burn dates that are seasonally grouped, even when they bridge the end of the calendar year.
Using 21 hotspot characteristics describing within-year seasonality, across-year return frequency, size and intensity, we produced global maps statistically discriminating the planet's most-different 10, 20, 50, 100, 500, 1000 and 3000 global Fire Regimes. Using principal component analysis to produce statistical Similarity Colors (see Accomplishment #3), we also visualized the degree of similarity among the different global Fire Regimes and graphically identified the fire characteristics responsible for the similarities and differences.
Geographically distant locations which share similar Fire Regime characteristics were found, including many Fire Regimes spanning across different hemispheres. Regularly occurring human-caused Fire Regimes, often associated with agricultural management, were also identified globally. Mirrored symmetrical latitude trend patterns are visible in each hemisphere, but latitude alone is insufficient alone to explain Global Fire Regime patterns. Pure, unblended primary statistical colors, which show within-year seasonality, are primarily found in temperate zones, but mixtures of primary colors are seen in the torrid zone, where fire seasonality is less marked. Fire Regimes having two distinct annual peaks or modes of fire frequency were the most common globally, followed by areas having three peaks per year. Bi-modal Fire Regimes typically have fire occurrence peaks both before and after the growing or monsoon season.
Locations sharing similar Global Fire Regimes have similar ecological effects and impacts from fire, and show where similar management knowledge and successful adaptation strategies might be borrowed, shared, or adopted. The date transform developed here to compare fire phenology globally can also be used to compare the phenology of plants, animals or other phenomena globally across hemispheres. Initial Global Fire Regime results were presented at US-IALE and AGU in 2014, and modified algorithm results were presented at US-IALE and AGU in 2018, and at ESA in 2019.
Global ecological studies based on observations are uncommon, yet the Research Ecologist has six examples of planetwide observation-based studies (Global Phenoregions for climatic change (Publication #56, 179 citations); Global Aquatic Ecoregions (Publication #76); Global TNC current and future terrestrial ecoregions (Publication #57, 117 citations), ForeCASTS, extrapolated from FIA (Publications #80, #90, #94, #100, #104); FLUXNET2015, based on global flux towers (“Publication” #108); Invasibility Zones, based on GloNAF; and Global Fire Regimes, based on MODIS hotspots), while ForWarn II is continental in scale.
Assessment of Severe Weather Damage to Forests
In April 2011, not long after its inception, ForWarn identified multiple damage tracks from an outbreak of tornadoes across the Southeastern United States. Damage tracks vary significantly in direction and width, and are not always recorded by the National Weather Service. ForWarn can monitor not only the initial damage but also the subsequent vegetation recovery following such storms. Our poster on detection and analysis of these tornadoes won the “Best Communication Product Award” at the 2012 International Users Conference. Using ForWarn in 2011, we were also able to see the large-scale simultaneous damage patterns from three hurricanes in the southeast, Rita, Ivan, and Katrina, and in 2018 the simultaneous damage tracks of both Florence and Michael.
With the launch and recent no-cost availability of 10m resolution data from the second Sentinel 2 satellite, operated by the European Space Agency, we have used these data to perform rapid, high-resolution forest damage assessments following 5 recent major hurricanes and tornadoes. Although not useful for finding new, unlocated forest disturbances, these higher resolution remote sensing assets are ideal for disturbances like hurricanes and tornadoes whose locations are already known. These rapid hurricane and tornado assessments employ the same custom forest disturbance algorithm that was developed for ForWarn (Publication #78, 68 citations, ), albeit used at the higher 10m Sentinel 2 resolution. We have made this new high-resolution Sentinel 2 imagery available within the ForWarn II online viewer, both as true color imagery, and as agricultural false color, which enhances vegetation disturbances. This ForWarn II online implementation is one of the first to enable simple, straightforward use of Sentinel 2 imagery by forest managers and non-specialists.
Hurricane Irma struck south Florida in September 2017, and the South Florida coastal mangroves had also been impacted earlier by Hurricane Katrina that passed over the peninsula in August 2005 before landing a second time in Louisiana. In addition to mapping the damage patterns, the NDVI multigraph tool in ForWarn II allowed direct comparison of mangrove recovery from the two storm systems. We mapped vegetation damage from Hurricane Maria that passed directly over Puerto Rico in September of 2017. Multi-date Sentinel 2 composites were important here in the tropics because of nearly perpetual cumulus clouds. Nevertheless, we presented a poster showing vegetation damage from Maria across the entire island at the 2018 US-IALE meeting. Florence's observable impacts were largely caused by flooding, given Florence's slow rate of spread inland. We mapped Florence damage in and near the Croatan NF in coastal North Carolina. Compared to Hurricane Michael, canopy damage was relatively isolated and of a lower degree.
The destructive forest impacts of Category 4 Hurricane Michael were captured by ForWarn II's routinely produced Early Detect product one week after the event, showing stark damage within a 50 km-wide swath stretching from the hurricane's track to the Apalachicola River. Using Sentinel 2 after the storm clouds parted, we were able to produce a detailed map of hurricane impacts, and even separate damage to evergreen and deciduous vegetation. Recently, we used ground-based severity observations, made by Karen Cummins and the Florida Forest Service, at more than 600 locations to classify a Sentinel 2-based damage departure map made using the ForWarn II algorithm, producing a wall-to-wall hurricane Michael damage severity map containing 6 different discrete damage classes. This is the first such hurricane damage severity map made using remotely sensed data, but classified with on-the-ground severity observations.
We are distributing these Rapid Storm Assessments to managers using the High-Resolution Forest Mapping (HiForM) website. In June 2019, the HiForM website received the 2019 Southern Research Station Director's Award for Excellence in Science Delivery (announcement, video, photo, ceremony, trophy). As we systematically improve Sentinel 2 rapid assessments using cloud computing via Google Earth Engine, the speed, quality and value of these new rapid vegetation damage capabilities increases with each new storm analysis. The Knoxville FIA office has begun coordinating with EFETAC on rapid damage assessments following storm events.
and Adjustment of Natural Phenological Timing, and Historical
Characterization of Dynamic Vegetation Behavior of Landscapes:
Subtle differences in phenological timing, when comparing current to past, are used in ForWarn for detecting forest disturbances (see Publication #78, 68 citations), but interannual differences in timing of phenology make direct comparisons of vegetation health and performance between years difficult, whether at the same or different locations. By "sliding" one phenology in time relative to the other, any particular phenological event can be synchronized across multiple locations, allowing direct comparison of timing differences in other phenological events. For example, using a linear sliding adjustment of time, one can synchronize the middle of the growing season across the nation. When the day counter reaches zero, it is the middle of the growing season across the entire CONUS. Now differences in the length of the growing season become obvious.
Time can also be "rubber-sheeted," compressed or dilated in non-linear ways. Considering one phenology curve to be a reference, time intervals within a second phenology curve can be "rubber-sheeted" to fit the first curve as well as possible, just by moving observations temporally in the second curve. Similar to georectifying a map inside a GIS, rubber sheeting a phenology curve also yields a warping signature that shows how many days the phenological development is ahead or behind the seasonality of the reference vegetation at every time interval. The size of these temporal shifts can quantify vegetation impacts from frost (dark yellow shows areas suffering frost delay), drought, wildfire, insects and diseases by permitting the most commensurate quantitative comparisons with unaffected vegetation.
The human calendar year, starting in January, is fixed, arbitrary and artificial, yet natural seasons differ significantly with location, beginning earlier in southern locations and aspects. Graphing NDVI greenness values in a circular polar plot, emphasizes the cyclical nature of annual phenology. Evergreen locations have a nearly circular annual plot, centered on the polar origin. But deciduous vegetation is off-center, shifted in the direction of the season of maximum greenness. Using polar vector statistics, we can calculate the degree of this shift, which then quantifies the degree of deciduousness of the vegetation growing at this location. This time of the year is also when the vegetation is at the peak of maximum greenness. The opposite direction across the annual cycle marks the period of minimum greenness, which is the mid-winter dormant or off-season. This dormant date, the period of least greenness activity, represents the most natural beginning/ending point for the annual phenological cycle at every location.
National-scale polar vector analysis of MODIS NDVI allows quantification of the degree of seasonality expressed by local vegetation across the map, and also selects the most optimum start/end of a local "Phenological Year" that is empirically defined by the vegetation that is growing at each location. Changes in the vegetation mixture or health status will be reflected in changes in the Phenological Year. The start and end of the Growing Season can be empirically defined as the day of year by which, say 15% and 85% of the total area under the circular greenness curve has been accumulated. Using six of these Polar Phenology Derivatives, we can produce a National Map of the 60 Most-Different Polar Phenoregions. These Polar Phenoregions tend to be more spatially cohesive than their non-polar counterparts, especially in the Pacific Northwest.
The spatio-temporal classification of annual MODIS NDVI into discrete, categorical phenoregions or "phenotypes" (see Publication #56, 179 citations) through time using MSTC (see Accomplishment #3) makes it possible for the first time to identify particular sequences of sequential annual phenotypes, thereby tracking multi-year trajectories of landscape change through time at any location. Examining the past behavior of all locations in a region or landscape permits an understanding of the diversity of behaviors it has shown, leading to insights into the flexibility or “brittleness” of that region. EFETAC has embraced this new approach as a potential method for quantifying the elusive idea of landscape resilience. Resilience is a desirable management goal, yet it is difficult to define, much less quantify. It is a scientific challenge to quantify the ecological resilience value of a local government's having spent hundreds of thousands of dollars on wildfire fuel thinning treatments, for example, as a justification for why the expensive process should continue in the next years.
Landscapes which have shown a broad diversity of behavioral trajectories in the past may be better able to cope with disturbances in the future, and are deemed more resilient. Quantifying the relative flexibility or brittleness seen at each location across a map, the Landscape Dynamics Assessment Tool (LanDAT) helps resource managers monitor broad patterns of historical vegetation change in order to gauge resilience and to understand their capacity to provide ecological services and benefits. A LanDAT Map Viewer, including polar NDVI plots for any location or time, lets users explore patterns for themselves using any browser.
LanDAT employs ideas from Information Theory to characterize past landscape vegetation dynamics. Because these ideas are so novel, these maps of Landscape Information Metrics have strange, unfamiliar names like Mutual Information and Conditional Entropy. But the highest-order maps summarizing Landscape Dynamics are Capacity, the diversity of past behaviors, Ascendency, the predictability of past behaviors, and Overhead, the un-predictability of past behaviors, all weighted by ecosystem productivity. LanDAT covers all land-use types, not just forests, and may form a quantitative foundation that can serve as the basis for a new paradigm of land management. The LanDAT Team has already held a number of workshops with interested land managers. . The spatial and temporal phenology classification “typing” from Multivariate Geographic Clustering (Accomplishment #3) provides the “phenotypes” upon which the LanDAT project is based.
Above-Ground Vertical Structure Types Using LiDAR, Virtual Mountain
LiDAR, which stands for Light Detection and Ranging, is a remote sensing method that uses a pulsed laser to measure variable distances through vegetation down to the Earth’s surface. LiDAR data contain information about the vertical distribution of above-ground vegetation biomass, but are voluminous, and present processing challenges. Usually collected for elevation information, forest structural data are often discarded or under-utilized when “shaving” the earth’s surface. Although of great potential value, forest managers face hurdles when trying to wrestle with LiDAR data. Needed are straightforward ways to process and convey the complex information contained in LiDAR surveys.
Following Hurricane Floyd in 1999, North Carolina became one of the first states to have comprehensive statewide LiDAR overflights. Portions of the State, however, were flown by six different subcontractors, each using LiDAR systems with different technical specifications. Due to data quality issues in eastern NC, we soon focused our LiDAR analyses on Western North Carolina, and the Great Smoky Mountains National Park (GSMNP). It is difficult to combine LiDAR datasets that differ in resolution, intensity and other characteristics.
We conceived a proportionalizing technique to “equalize” LiDAR data collections that differ in resolution or point density, allowing them to be used together in the same map. We group all LiDAR returns occurring within each map grid cell of, say, 30m, and calculate the proportion of the total returns in that cell coming from each 1m height “layer.” The percentage of total returns at every height interval sum to 100% up through the forest canopy, regardless of the initial LiDAR pulse resolution or total return counts per grid cell.
Then, we utilize the same Multivariate Geographic Clustering methods that we used to produce ecoregions, phenoregions and NEON domains (see Accomplishment #3). Using the supercomputers at ORNL, we clustered the list of percentages of total returns at every vertical level, upward through the canopy. This groups all locations together that show a similar vertical distribution profile of vegetation. A map is produced that shows categorical forest types which have the same profile shapes of vertical above-ground biomass distribution upwards through their canopies. Each clustered Forest Structural Type group that is statistically formed has a vertical frequency graph that shows the profile of the overstory, understory and shrub layers present in that type of forest.
The resulting Forest Vertical Structure Type Maps show as categorical colors all areas on the map sharing similar vertical profile distributions of aboveground biomass. Colored in statistical Similarity Colors (see Accomplishment #3), the maps show degrees of difference in vertical structure, and are seamless, with no artefacts or interruptions visible between areas collected in different LiDAR flights.
Using MapCurves (see Section 3C, Acomplishment #2, and Publication #59, 88 citations), we overlaid the map of clustered GSMNP Forest Vertical Structure Types with the best vegetation type map for the Park, to obtain a translation table between vertical structure and forest species composition. A many-to-one or many-to-many crosswalk relationship existed between Forest Vertical Structure Types and Forest Vegetation Types within the Park. Some Vertical Structure Types that corresponded to the same Forest Type may be showing increasing development of overstory height and understory with increasing stand age. Thus, there is a possibility that one could produce a map estimating the age of forest stands using clustered LiDAR.
Robert Whittaker famously used plots in GSMNP to demonstrate his gradient analysis hypotheses, showing the continuous relationships between types of forest vegetation and environmental gradients of moisture and elevation. To summarize the continuous relationships of vegetation to elevation and aspect, we conceptualized a single, imaginary conical Virtual Mountain, depicted as a circle whose peak is at the center. We located every 30m grid cell physical location within the Park onto this single Virtual Mountain, using its elevation and aspect as coordinates. Because many Park locations share the same aspect and elevation characteristics, they are all located at the same point on the Virtual Mountain, even if actually separated by large geographic distances. Virtual Mountain plots represent the ultimate Whittaker gradient analysis, since they exhaustively utilize every 30m cell contained within the GSMNP.
We show multiple Virtual Mountain maps for each LiDAR-derived variable (canopy height, for example), one mapping the mean with aspect and elevation, and others depicting the variability across the population of Park locations at that elevation and aspect. Clustered Vertical Structure Types, shown on a Virtual Mountain, show subtle and interpretable relationships with aspect and elevation. Forest aspect and elevation relationships shown in Virtual Mountain presentations are easily and intuitively understood by resource managers, and we believe the GSMNP LiDAR relationships are representative of gradient relationships throughout most forested areas in Western North Carolina.
Our LiDAR maps, Virtual Mountain plots, and LiDAR profile clusters from our 2015 paper (Publication #101) were presented at the GSMNP Science Colloquium in 2016, and are available in an ORNL DAAC database, DOI 10.334/ORNLDAAC/1286, that was updated in 2018 to include both the TN and NC sides of the Park seamlessly, along with maximum canopy height maps. Maps of clustered Forest Vertical Profile Types and Virtual Mountain Plots represent two new ways to make complex LiDAR data more accessible to resource managers.