AUTHORS: Geoffrey M. Henebry*, South Dakota State University
ABSTRACT: In our current era of abundant medium spatial resolution (10m-60m) multispectral data, it is a good time to take stock of change analysis techniques that are well suited to temporally dense subannual data. Here I review some helpful concepts about change analysis and theory about the seasonal development of spatial pattern. I pull examples of spatial-temporal change detection from earlier literature as well as provide a worked example of comparative landscape dynamics using USGS’s new Landsat Analysis Ready Data (ARD).
AUTHORS: Lan H. Nguyen, Geospatial Sciences Center of Excellence, South Dakota State University; Geoffrey M. Henebry, Geospatial Sciences Center of Excellence and Department of Natural Resource Management, South Dakota State University
ABSTRACT: Despite its low population, the Northern Great Plains (NGP) region of the United States of America has been undergoing substantial land cover/land use change (LCLUC) over the past two decades, including conversion of grassland to cropland, urbanization, fossil fuel mining, and expansion of irrigation. Better detection of rapid land change is critical for management and conservation of diminishing prairie habitats and ecosystem services and for projecting crop and biofuels production and their impacts on rural transportation infrastructure. Most change detection strategies that have been commonly used in remote sensing studies were developed in an era of image scarcity and limited computation power, and thus focus on comparing just a few scenes. In the current era of abundant earth observations, the better approach takes advantage of all available data. Here, we demonstrate an approach to detect LCLUC using land surface phenology. Our study area spans multiple regions of interest across the NGP where land use transitions have been documented: near ethanol bio-refineries, major towns and cities, and fossil fuel mining sites. First, we compute accumulated growing degree-days (AGDD) from MODIS 8-day land surface temperature. Using EVI derived from Landsat surface reflectance, we then fit at each pixel a downward convex quadratic model to each year’s progression of AGDD (EVI = a+ß×AGDD-?×AGDD2). We then use information from the fitted model as input to two different classification techniques, e.g., linear mixture analysis (LMA) and random forest classifiers (RFC), to generate a land use/land cover map of the study area. We use the USDA Cropland Data Layer (CDL) as a reference for LMA (choose endmembers) and RFC (select training and validation dataset), while acknowledging the CDL’s limitations. By tracking the proportions (LMA output) or land cover map (RFC output) over the 2001-2011 period, we characterize and quantify various land change transitions across the NGP.
AUTHORS: Ranjeet John*, Jiquan Chen, Wenjuan Ma, Hogeun Park, Sarah Hession – Michigan State University, USA; Vincenzo Giannico, The University of Bari Aldo Moro, Italy
ABSTRACT: Rapid land cover/use (LCLU) has resulted in the degradation of structure and function of temperate, semiarid grassland ecosystems on the Mongolian Plateau (MP). The MP is a complex socio-ecological system which has undergone substantial changes following the post-1970s and post-1990s liberal reforms in Inner Mongolia and Mongolia, as well an increased frequency of drought and extreme winters. Mongolia and Inner Mongolia have similar ecological climatology and biodiversity, but have divergent trends owing to differences in political regimes and land management policies. Here we use the coupled natural human (CNH) system framework to explain cause and effect between the spatial and temporal variability of above ground biomass (AGB) across a precipitation gradient with trends in extreme climatic events as well as socioeconomic drivers. We employed structural equation modeling (SEM) to integrate long term data records derived from MODIS (2000-2016) and GIMMS3g NDVI (1981-2015) which serve as proxies of productivity, MERRA-2 reanalysis climate data, and socioeconomic drivers at the county level that include total livestock density and population density and distance to towns and cities as a proxy of access to social goods and services. Our sampling units were stratified by the dominant steppe types, namely meadow, typical and desert steppe. We used SEM and geographically weighed regression (GWR) to explain AGB biomass degradation as a function of migration trends in response to push/pull factors including extreme events such as drought and global financial crisis. In this study, we seek to isolate the effects of human influence on AGB from the strong precipitation-driven trends on the MP. The AGB response to a combination of anthropogenic modification and extreme climate events as modeled by SEM and GWR showed significant spatio-temporal variability which was explained by extreme climate events, distance to urban/built-up areas and livestock density.
AUTHORS: Steve Norman*, William Hargrove – US Forest Service
ABSTRACT: In deciduous forests, remote sensing-based assessments of fire impacts are often confounded by the seasonal flux of land surface phenology (LSP). In the Southern Appalachians, USA, fires burn primarily in late winter/early spring and during or immediately after leaf abscission in the fall. This timing can make it challenging to recognize fire impacts apart from ongoing phenological transitions, particularly in cloudy regions where clear imagery is infrequent. More fundamentally, wildfire impacts, like those of low intensity prescribed fire, are often structurally isolated to just the understory, leaving overstories immeasurably affected. While such structurally selective impacts can have huge implications for wildlife and fuels management, the independent phenological detectability of these elements is routinely ignored. In this research, we use 10m seasonal maximum NDVI composites in both the growing season and winter to isolate deciduous and evergreen impacts separately for numerous fall 2016 wildfires that burned NC, SC, GA and TN. We then compare differences in season-specific impacts across topographic gradients. Results show greater winter than summer impacts, but spatial coherencies are consistent with fire behavior and impacts observed in the field. Insights suggest the value and need for formal incorporation of LSP into wildfire and forest monitoring and a reappraisal of how we conceptualize fire severity in this widespread forest type.
AUTHORS: Morgan A. Crowley*, Jeffrey A. Cardille – McGill University; Michael A. Wulder, Joanne C. White – Canadian Forest Service (Pacific Forestry Centre)
ABSTRACT: The 2017 fire season was the largest on record for British Columbia (BC) in terms of area burned, and mapping these fires is important for monitoring forest-disturbance impacts. Two data sources provide reliable estimates of post-fire perimeters: (1) the Composite-to-Change (C2C) protocol, which makes annual forest disturbance mapping possible for Canada using annual best-available-pixel composites; and (2) the Canadian National Fire Database (CFDB), which identifies fire perimeters using data from fire management agencies. These resources are vital records of fire locations and extents, but are not intended to map active fire growth and patterns through time. The purpose of this project is to determine how these data sources can inform mapping the fire’s active phase. In addition to the observations provided by Landsat, imagery from sensors such as Sentinel-2, ASTER, and MODIS can be used to increase the temporal resolution and data density. We used the Bayesian Updating of Land Cover Classifications (BULC) algorithm to merge initial classifications while tolerating occasional smoke and clouds. Working in areas identified by the CFDB, we compared each source data’s post-fire Normalized Burn Ratio (NBR) against the expected NBR of a pre-fire best-available-pixel image from C2C. This approach enabled the classification into fire/no-fire using the large-scale processing power of Google Earth Engine. BULC then fused these classified images in Earth Engine, producing a series of updated fire extent maps for the 2017 fire season in BC. The increased temporal density allows us to retrospectively synthesize and, analyze the dynamics and growth of individual fire events, which was not possible using previously established methods. This capacity can support reconstructing the progression of active fire lifespans to better understand fire growth and underlying drivers. Methods from the case study could be scaled-up for nation-wide fire mapping at intermediate time-steps for past and future fire seasons.
AUTHORS: Forrest M. Hoffman, Oak Ridge National Laboratory; Zachary L. Langford, University of Tennessee Knoxville; Jitendra Kumar, Oak Ridge National Laboratory; Steve Norman, USDA Forest Service Eastern Forest Environmental Threat Assessment Center (EFETAC); William W. Hargrove, USDA Forest Service Eastern Forest Environmental Threat Assessment Center (EFETAC)
ABSTRACT: The Arctic landscape is changing at unprecedented rates due to climate change, natural disturbances, and anthropogenic development. Moreover, the Arctic is also projected to warm at a rate twice that of the global average in the coming century, accelerating climate feedbacks, including those from wildfire disturbances. Wildfires are the dominant disturbance impacting boreal forests, and climate-driven shifts in fire frequency, intensity, and extent produce large changes in fire regimes. To understand changing fire regimes, we applied a random forest machine learning algorithm to estimate wildfire locations based on 500-m resoluton MODIS and 1-km resolution Daymet variables in the State of Alaska. We employed the Google Earth Engine (GEE) platform and its inventory of satellite imagery, meteorology, elevation, and atmospheric data to detect and map the 70 large boreal fires of the 2004 Alaska wildfire season. Using the national Monitoring Trends in Burn Severity (MTBS) data for validation, our random forest method achieved scores of 0.87, 0.95, and 0.91 for the precision, recall, and F1-score metrics when multiple vegetation indices, elevation, and individual spectral bands were incoporated into the analysis. Inclusion of Daymet data with no time lags did not significantly improve the performance of the detection method. The EVI and SAVI indices, plus the NIR spectral band, were the most significant inputs to the random forest algorithm. Results suggest that large burned areas may not be reported and included in the MTBS database. In future work, we plan to extend this methodology to higher resolution observations (15 m to 30 m) from Landsat.
AUTHORS: Jitendra Kumar, Oak Ridge National Laboratory, Oak Ridge, TN; William W. Hargrove, USDA Forest Service, Asheville, NC; Steven P. Norman, USDA Forest Service, Asheville, NC; Forrest M. Hoffman, Oak Ridge National Laboratory, Oak Ridge, TN
ABSTRACT: Fire Regimes are conceptually useful to land managers and are qualitatively understood, but few quantitative techniques exist for empirically delineating geographic regions whose wildfire spatial and temporal characteristics, re-visitation frequency and intensities are similar. We present a comprehensive effort which considers the extensive and consistent thermal “hotspot” data which are collected globally by the two MODIS sensors during their 17-year orbital history. Such consistent and ubiquitous remote sensing data provides an opportunity to produce a nearly two-decade quantitative discrimination of different global fire regimes, including tele-connections across hemispheres. We do not filter or remove human-caused fires from wildfires, instead considering and classifying both types of fire regimes holistically. To appropriately address opposing seasonal juxtaposition across northern and southern hemispheres we develop a special transformation of fire dates which allows statistical identification and discrimination of, say, “summer” fires, regardless of the calendar month in which they occurred across the hemispheres. This date transform permits the delineation of similar fire regimes which occur in both the northern and southern hemispheres, without causing any discontinuities at the equator. On the basis of about 20 descriptive “hotspot” variables, we produced a series of global maps at multiple levels of fire regime discrimination. By applying principal component analysis on the 20 “hotspots” variables we also quantify the degree of similarities among the different global fire regimes and quantitatively identify the reasons, or characteristics, for the similarity or the differences. Several examples of geographically distant locations which share similar fire regime characteristics were found, and some of these fire “tele-connections” span across different hemispheres. Locations experiencing the same or similar global fire regimes may have similar ecological effects and impacts from fire, and similar management knowledge and successful adaptation strategies might be borrowed, shared, or adopted. Regularly occurring human-caused fires can also be easily identified globally.
AUTHORS: Richard Tran Mills, Argonne National Laboratory; Vamsi Sripathi, Intel Corporation; Sarat Sreepathi, Oak Ridge National Laboratory; Forrest M. Hoffman, Oak Ridge National Laboratory; William W. Hargrove, USDA Forest Service Southern Research Station
ABSTRACT: The increasing availability of high-resolution geospatiotemporal data sets from sources such as observatory networks, remote sensing platforms, and computational Earth system models has opened new possibilities for knowledge discovery using data sets fused from disparate sources. Traditional algorithms and computing platforms may be impractical for the analysis and synthesis of data sets of this size; however, new algorithmic approaches and implementations that can effectively utilize the complex memory hierarchies and the extremely high levels of available parallelism in state-of-the-art high-performance computing platforms can enable such analysis. We describe a hybrid parallelism (MPI-OpenMP) based implementation of accelerated k-means clustering and some optimizations to boost computational intensity and utilization of wide SIMD lanes and many hardware threads on state-of-the art multi- and manycore processors, including the second-generation Intel Xeon Phi ("Knights Landing") processor based on the Intel Many Integrated Core (MIC) architecture, and on cutting edge GPGPU architectures, and we explore several applications thereof to large-scale analysis of MODIS NDVI and LiDAR-derived forest ecosystem data sets.
ABSTRACT: Remote sensing approaches designed to spatially and temporally mapping seasonal inundation provide critical information for many applications which include water accounting, flood control and forecast and groundwater recharge estimation. Deep-learning algorithms have recently become a hotspot in the geoscience and remote sensing. In addition, as the most representative supervised DL model, convolutional neural network(CNN) have outperformed most algorithms in many fields. In this study, a framework of CNN for seasonal inundation mapping would be presented for South China-a key area of China historically, politically and economically. Available 8-day MODIS data from the year of 2000 would be used here. Firstly we labeled the fraction of water cover of a few training samples using higher resolution RS data. Then a convolutional neural network is trained with Tensorflow to generate a time-series of the fraction of water cover for Southeast China. Each pixel would be classified by annual and seasonal fraction of water cover for further analysis with physical and social data. Furthermore, the cross-validation would be implemented. And ground water flow observation would be also used for validation. The results of this study demonstrate the potential for the use of MODIS data for temporal and spatial detection of water inundation and help to understand the change of water inundation in regional scale from 2000. The impact of climate change and human activity on water inundation would be discussed here as well.
AUTHORS: Nicholas Kruskamp*, Joshua Gray, Ross Meentemeyer – North Carolina State University
ABSTRACT: Accurate maps of host plant species are essential for predicting disease spread in forest environments, but species-level vegetation data rarely exist at sufficiently high accuracies or spatial resolutions to model epidemiological processes. We examine the manner in which host map accuracy and resolution affect spatio-temporal models of disease spread. Sudden oak death is used as a case study of an emerging infectious disease requiring predictive models with reliable host maps for disease management and quarantine policy. We used field measurements, muli-temporal moderate resolution Sentinel-2 imagery, and a random forest algorithm to map the density and biomass of tanoak (Lithocarpus densiflorus) - the reservoir host species undergoing substantial tree mortality at a new outbreak of international concern in southwest Oregon. Moderate-resolution remotely sensed images from Sentinel-2 provided spectrally-derived environmental covariates as predictor variables in the random forest model. Seventy-two field sites were surveyed for species density and composition. Elevation, G(reen)NDVI, solar illumination, and the first image principal component were the most important variables in the random forest model. Our mapping approach resulted in significantly less area of the landscape occupied by tanoak compared to existing vegetation datasets (CALVEG, LEMMA). Preliminary results suggest that the more accurate, highly resolved maps of tanoak density have a major impact on epidemiological modeling results, with slower rates of disease spread caused by the reduced extent of the host species. Our results have substantial implications for disease management. Results from this study suggest that remote sensing with machine learning techniques should be more commonly used to improve landscape epidemiological modeling inputs, and more critically examine the impact of input host species maps on model outputs.
AUTHORS: William Hargrove*, Steve Norman – USDA Forest Service; Jitendra Kumar, Forrest Hoffman – Oak Ridge National Laboratory
ABSTRACT: 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" the two phenologies in time using a Procrustean linear time shift, any particular phenological event or "completion milestone" can be synchronized, allowing direct comparison of differences in timing of other remaining milestones. Going beyond a simple linear translation, time can be "rubber-sheeted," compressed or dilated. Considering one phenology curve to be a reference, the second phenology can be "rubber-sheeted" to fit that baseline as well as possible by stretching or shrinking time to match multiple control points, which can be any recognizable phenological events or dates. Similar to "rubber sheeting" to georectify a map inside a GIS, rubber sheeting a phenology curve also yields a warping signature that shows at every time and location how many days the adjusted phenology is ahead or behind the phenological development of the reference vegetation. Using such temporal methods to "adjust" phenologies may help to quantify vegetation impacts from frost, drought, wildfire, insects and diseases by permitting the most commensurate quantitative comparisons with unaffected vegetation.
AUTHORS: Matthew Garcia*, Philip A. Townsend – Department of Forest and Wildlife Ecology, University of Wisconsin-Madison; Brian R. Sturtevant, Institute for Applied Ecosystem Studies, Northern Research Station, USDA Forest Service
ABSTRACT: We developed novel methods for processing dense Landsat time series over the 1984-2013 period for the forests of the western Lake Superior region, USA. These methods allow us to derive mean annual phenology, to attribute deviations from that mean phenology to interannual variability in climatological conditions, and to examine further deviations from the expected phenoclimatology attributable to disturbance. The heavily-forested area of northeastern Minnesota along Lake Superior has seen numerous disturbance events within our study period: large and small fires, severe wind events, drought periods, and both focused and widespread insect outbreaks. While the more severe disturbances are often easier to identify and attribute to known causes, potentially mild and often slow-building events such as insect defoliation and drought stress can be more challenging to assess from the remote sensing perspective. We have documented observations of outbreaks of two defoliating insects, the eastern spruce budworm (Choristoneura fumiferana) and the forest tent caterpillar (Malacosoma disstria), within our study area and period. Outbreak events for both of these agents were identified in the Landsat image record and led to clear departures from typical or expected interannual variability in their respective host phenology. These disturbances can be tracked from the initial event through the forest regeneration period, helping us better understand the cycles of forest disturbance and recovery that result from quasi-periodic defoliating insect outbreaks that are comparatively mild relative to stand-replacing disturbances such as fire or logging.
AUTHORS: Zhiyue Xia, Department of Ecosystem Science and Management, The Pennsylvania State University; Laura Leites, Department of Ecosystem Science and Management, The Pennsylvania State University; Douglas Miller, Department of Geography and Department of Ecosystem Science and Management, The Pennsylvania State University; Andrew Liebhold, US Forest Service Northern Research Station, Morgantown, WV
ABSTRACT: The Gypsy moth, Lymantria dispar, is an exotic forest pest that was introduced to the USA in 1869. Since then it has spread continuously across the majority of the northeastern US. Larvae of this insect prefer feeding on oak species, although other species also serve as host trees. During outbreak, larvae defoliate forests across large regions and repeated defoliation can predispose the trees to attacks by secondary insect pests or fungal infections causing tree mortality. Gypsy moth outbreaks are episodic and are difficult to predict. Development of forecasting models remains a challenge despite their potential usefulness in effectively mobilizing resources to deal with the outbreaks. Previous studies indicate that vegetation attributes measured through remote sensing, as well as terrain, and climate characteristics influence the likelihood gypsy moth outbreaks. In addition, temporal and spatial variables describing the cyclic and spatial patterns of the outbreaks could be very valuable in forecasting outbreaks.In this study, we develop a model that forecasts gypsy moth outbreaks using Pennsylvania as a case study. We use systematic sampling and locate 8,424 sampling points across forest areas of Pennsylvania and focus on outbreaks during the time period 200-2016 to train the model. For each point, a large suite of temporal and spatial predictor variables are derived from remote sensing, climate, topographic and inventory data, while the occurrence of the outbreak is obtained from annual outbreak sketch maps. We use the machine learning modeling algorithm Random Forests which has a well-documented predictive ability and can deal with a large number of variables. We present modeling results and an assessment of the model performance that focuses on hindcasting outbreaks during the period 1980-1990. An accurate forecasting model will be of critical importance for projecting the spatial extent of future outbreaks and for forest management planning.
