Spatial Analysis For Conservation and Sustainability
Remote Sensing
Satellite images provide a wonderful record of the last fifty years of global change. We have pioneered new methods to map wildlife habitat and proxies for biodiversity and habitat, as well as agricultural abandonment and other types of land use change for large areas. We analyze MODIS/VIIRS data across the globe, Landsat and Sentinel-2 across continents, and high-resolution CORONA spy satellite imagery across countries.
Background: Wildland–urban interface (WUI) maps identify areas with wildfire risk, but they are often outdated owing to the lack of building data. Convolutional neural networks (CNNs) can extract building locations from remote sensing data, but their accuracy in WUI areas is unknown. Additionally, CNNs are computationally intensive and technically complex, making them challenging for end-users, such as those who use or create WUI maps, to apply.
Aims: We identified buildings pre- and post-wildfire and estimated building destruction for three California wildfires: Camp, Tubbs and Woolsey.
Methods: We evaluated a CNN-based building dataset and a CNN model from a separate commercial vendor to detect buildings from high-resolution imagery. This dataset and model represent to end-users the state of the art of what is readily available for potential WUI mapping.
Key results: We found moderate accuracies for the building dataset and the CNN model and a severe underestimation of buildings and their destruction rates where trees occluded buildings. The CNN model performed best post-fire with accuracies ≥73%.
Conclusions: Existing CNNs may be used with moderate accuracy for identifying individual buildings post-fire and mapping the extent of the WUI. The implications are, however, that CNNs are too inaccurate for post-fire damage assessments or building counts in the WUI.
Accurate maps of gains in tree cover are necessary to quantify carbon storage, wildlife habitat, and land use changes. Satellite-based mapping of emerging smallholder woodlots in heterogeneous landscapes of sub-Saharan Africa is challenging. Our goal was to evaluate the use of time series to detect and map small woodlots (<1 ha) in Tanzania. We distinguished woodlots from other land cover types by woodlots’ distinct multi-year spectral time series. Woodlots exhibit greening from planting to maturity followed by browning at harvest. We compared two time series approaches: 1) a linear model of Tasseled Cap Wetness (TCW) and other indices, and 2) LandTrendr temporal segmentation metrics. The approaches had equivalent woodlot detection accuracy, but LandTrendr segments had lower accuracy for characterizing woodlot age. We tested the effect of the following factors on woodlot detection and mapping accuracy: the length of the time series (2009–2019), frequency of observations (all Landsat vs. only Landsat-8), spatial resolution (30-m Landsat vs. 10-m Sentinel-2), and woodlot age and size. Woodlot mapping accuracies were higher with longer time series (54% at 3-yrs vs 77% at 7-yrs). The accuracies also improved with more observations, especially when the time series was short (3-yrs Landsat-8 only: 54% vs. all-Landsat: 64%, p-value <0.001). Sentinel-2’s higher spatial resolution minimized commission errors even for short time series. Finally, less than half of young and small (<0.4 ha) woodlots were detected, suggesting considerable omission errors in our and other woodlot maps. Our results suggest that the accurate detection of woodlots is possible by analyzing multi-year time series of Landsat and Sentinel-2 data. Given the region’s woodlot boom, accurate maps are needed to better quantify woodlots’ contribution to carbon sequestration, livelihoods enhancement, and landscape management.
Detailed maps of forest structure attributes are crucial for sustainable forest management, conservation, and forest ecosystem science at the landscape level. Mapping the structure of broad heterogeneous forests is challenging, but the integration of extensive field inventory plots with wall-to-wall metrics derived from synthetic aperture radar (SAR) and optical remote sensing offers a potential solution. Our goal was to map forest structure attributes (diameter at breast height, basal area, mean height, dominant height, wood volume and canopy cover) at 30-m resolution across the diverse 463,000 km2 of native forests of Argentina based on SAR Sentinel-1, vegetation metrics from Sentinel-2 and geographic coordinates. We modelled the forest structure attributes based on the latest national forest inventory, generated uncertainty maps, quantified the contribution of the predictors, and compared our height predictions with those from GEDI (Global Ecosystem Dynamics Investigation) and GFCH (Global Forest Canopy Height). We analyzed 3788 forest inventory plots (1000 m2 each) from Argentina’s Second Native Forest Inventory (2015–2020) to develop predictive random forest regression models. From Sentinel-1, we included both VV (vertical transmitted and received) and VH (vertical transmitted and horizontal received) polarizations and calculated 1st and 2nd order textures within 3 × 3 pixels to match the size of the inventory plots. For Sentinel-2, we derived EVI (enhanced vegetation index), calculated DHIs (dynamic habitat indices (annual cumulative, minimum and variation) and the EVI median, then generated 1st and 2nd order textures within 3 × 3 pixels of these variables. Our models including metrics from Sentinel-1 and 2, plus latitude and longitude predicted forest structure attributes well with root mean square errors (RMSE) ranging from 23.8% to 70.3%. Mean and dominant height models had notably good performance presenting relatively low RMSE (24.5% and 23.8%, respectively). Metrics from VH polarization and longitude were overall the most important predictors, but optimal predictors differed among the different forest structure attributes. Height predictions (r = 0.89 and 0.85) outperformed those from GEDI (r = 0.81) and the GFCH (r = 0.66), suggesting that SAR Sentinel-1, DHIs from Sentinel-2 plus geographic coordinates provide great opportunities to map multiple forest structure attributes for large areas. Based on our models, we generated spatially-explicit maps of multiple forest structure attributes as well as uncertainty maps at 30-m spatial resolution for all Argentina’s native forest areas in support of forest management and conservation planning across the country.
Habitat selection is a fundamental behavior of species that shapes a wide range of ecological processes, including species distribution, abundance, nutrient transfer, and tropic dynamics. The study of habitat selection is important to understand the interaction between species and environment. But it is a multivariate and hierarchical process, in which species are distinctively affected by several factors at multiple spatial scales. Therefore, it is important to understand how species select their habitat, what are the important spatial scales, and how the habitat selection process varies for different species.
Figure 1: Conceptual hierarchy of the decision-making process of habitat use by a migratory songbird (Stanley et al 2021).
Hierarchical habitat selection in birds varies greatly by species due to their ecological niches and behaviors. For instance, the Northern Spotted Owl specializes in old-growth conifers for nesting, forages in mature forests, and prefers undisturbed landscapes for its home range. Conversely, the Kirtland’s Warbler prefers, early to mid-successional jack pine forests, growing on sandy soil for nesting, these forests provide the specific vegetation structure and insect abundance that are essential for their foraging needs. Studying habitat selection is therefore crucial for effective conservation and ecosystem management, as it provides insights into their ecological requirements and aids in preserving their populations and the overall health of ecosystems.
Figure 2: Kirtland’s Warbler (left) Spotted Owl (right)
Despite notable advancements in the field, our understanding of the hierarchical aspects of habitat selection in birds remains limited. Habitat selection models typically rely on satellite data from a single sensor and scale, which limits their effectiveness in capturing spatial patterns of bird habitat.
Akash Anand is currently conducting a study aimed at modeling multiscale hierarchical habitat selection in birds and explaining the factors influencing individual species’ choices. His research investigates the crucial spatial scales for different species and identifies local environmental features that play a pivotal role in overall habitat selection decisions. To achieve this, he employs deep learning models to gain insights into the intricate interactions between species and their environments.
In conclusion, Akash’s research aims to determine the crucial spatial scales for individual species, providing valuable insights for conservationists and policymakers. Additionally, the findings will provide evidence of how the same species respond to varying environmental conditions and how their choices differ in different scenarios. This knowledge will inform more effective conservation and management strategies.
Halting biodiversity declines and promoting sustainable ecosystem usage are major conservation goals. To do so, it is necessary to understand the environmental correlates of biodiversity patterns.
Environmental variables used in biodiversity modelling come from a variety of sources and have varying levels of power to explain distributions of different species. Many environmental variables that have been used regularly for many years have shortcomings- they may not cover large areas, may not capture suitable habitat, or may not be able to capture changes in environmental conditions over time or space. Increasingly, novel remotely-sensed environmental data are being developed for modelling biodiversity patterns. Novel remotely sensed products may complement or even offer better results than environmental variables that have been used for many years.
Olah set out to identify sets of complementary variables, from among a set of standard variables and newly created variables, that can improve species distribution modelling. Olah used a combination of land cover, elevation, precipitation, and temperature variables, that are commonly used in species distribution modelling, and a set of novel-remotely sensed products to model distributions of forest affiliated bird species in Argentina.
Yungas forest in Calilegua National Park, Jujuy, Argentina
The set of novel environmental variables were created by SILVIS lab postdoctoral researcher Eduarda Silveira. These products measure spatial and interannual variation in the phenology of land surface temperature and forest vegetation greenness. Olah predicted that areas with more spatial variability in phenology and thermal conditions are more likely to host more species because there are a variety of resources and thermal conditions in close proximity, allowing many species to coexist in a small area. These areas may also buffer against high year-to-year variation in conditions because organisms are more likely to have access to refugia or resources that could allow them to persist. Temporal variation in forest greenness or temperature describes how consistent conditions are between years. High variability means that phenological events are not occurring at a predictable time, while low variability means that events are occurring predictably each year.
In another new product developed by Silveira, ground forest inventory data was combined with radar-based remotely-sensed data, resulting in modelled forest structure wall-to-wall across Argentina. Silveira also developed maps of forest phenoclusters and phenocluster diversity. Phenoclusters classify different forest types in Argentina, based on vegetation phenology, land surface temperature, and precipitation. Olah thinks phenoclusters are a more ecologically relevant way to characterize habitat important to bird species than typical land cover maps. Phenoclusters capture functional rather than only compositional or structural characteristics. By comparing how well these novel remotely-sensed products and traditionally used variables perform in species distribution modelling Olah assessed their usefulness for biodiversity mapping.
Polylepis forest in an Andean Mountain valley, Jujuy, Argentina.
Olah developed species distribution models for 152 forest bird species. She found that among three sets of models she constructed, those containing novel, traditional, or a mixed set of variables, performance was similar. However, models constructed from the mixed set of variables performed slightly better than models containing only one or the other set of data. The variables that were included in the greatest number of individual species’ distribution models included precipitation seasonality, precipitation of the driest quarter, as well as spatial heterogeneity in winter land surface temperature, which is a novel variable. Her results highlight how variables derived from different sources can offer complementary information for biodiversity modelling. Her models contribute to forest harvest planning in Argentina.
Long-distance migrations are an important part of the life cycle for most Arctic waterfowl. The birds spend several months each year between their breeding and wintering places, with most of this time staying at their stopover sites to rest and feed. Habitat quality at the stopovers determine subsequent survival during migrations and reproductive success. Two questions present themselves: where are the key waterfowl habitats along their migratory routes, and what landscape features make these places vitally important? Understanding this is crucial for bird protection and population management.
Fig. 1 Bewick’s Swan
Satellite images are used widely for landscape research; however, their application to bird migration studies is challenging. First, environmental conditions within a large area are not the same during migration period, so there is no single time window suitable to select satellite images for the entire area. Second, favorable conditions at each stopover site are short-term, as birds spend only a few weeks at each site. The needed satellite images may be unavailable for those exact periods due to weather conditions (clouds etc.), making it difficult to obtain sufficient data for a single migratory season. Third, environmental conditions and migration time could vary by several weeks depending on weather of a given year, which, in turn, impedes combining images from different years.
Natalia studied how to use satellite images for delineating key migratory habitats using the example of spring migration of Bewick’s Swan in Eurasian steppe (fig. 1). To overcome the limitation related to satellite image availability, Natalia used snow melt as a synchronizing point. In practice, specific dates on which the swans appear at one or another area do not really matter because the birds will move farther north as soon as ice-free water and food become available there. In other words, they arrive soon after snow melt and stay a couple of weeks until more northern areas become snow-free.
Natalia used daily MODIS data to identify where snow melted each year in different parts of the swan’s migratory flyways and then filtered Landsat images using that information. This allowed her to combine Landsat images for all migration areas from different years at the same phenological phase. This approach made it possible to produce accurate and detailed landscape maps demonstrating what environmental conditions prevailed at stopovers at exactly the time when the swans were present. With these maps species distribution modeling can delineate key swan’s habitats.
Fig.2 Key habitats during summer
The maps have revealed the critical landscape feature important to the birds: numerous local depressions scattered across croplands (fig. 2). In summer these appear as a part of the agricultural landscape (only the smallest, lowermost places may not be ploughed and get overgrown with wildflowers) and are hardly detectable (fig. 3, e) on satellite images. In spring, however, they accumulate melted water to become shallow temporary water bodies (fig. 3, a-c). These flooded depressions provide migratory birds with food and refuge so that the waterfowl do not need to move between roosting and feeding sites. They are also available 10-15 days before ice-out on lakes, allowing birds to migrate and potentially reach breeding grounds earlier.
Fig.3 Changes in open water in the steppe during spring. RGB: SWIR 1, NIR, Red bands. Red circle –flooded fields, yellow square -permanent lake. Bright blue color indicates snow/ice, black color indicates open water.
More generally, Natalia’s research provides a useful approach to understanding the key short-term conditions that birds rely on during migrations. Natalia’s results have important implications for conservation efforts, such as the creation of protected areas and free hunting zones and adjusting land management in agricultural lands.
The boreal forests of Russia, Canada, Alaska, and Scandinavia, collectively known as the circumboreal region, contain 25% of global forests and 50% of the world’s soil carbon pool. Wildfire is an important disturbance agent within the circumboreal region and plays a key role in facilitating its nutrients cycling and succession. Climate change is warming the circumboreal region at twice the rate of the global average creating longer, hotter, and drier wildfire seasons. This in turn has warmed and dried the circumboreal growing season, stressed vegetation, and may have contributed to the severe 2019-2021 wildfire seasons. The specter of increasing burned area extent within the circumboreal region is concerning because it may volatilize the region’s enormous soil carbon pool, exacerbating climate warming.
Ground-based mapping of burned area within the circumboreal region is challenging due to its remote and inaccessible nature. Instead, burned area is typically mapped using satellite-based remotely sensed datasets. Mapping burned area within the Eurasian boreal region prior to the year 2000 is difficult due to a lack of available Landsat and MODIS data. Only the coarse resolution AVHRR archive provides data capable of exhaustively mapping wildfire throughout the circumboreal region over multi-decadal timescales. Mapping wildfire with AVHRR data is complicated by sensor-based limitations that, if unaccounted for, will degrade the quality of resulting burned area maps. An advantage of statistical models over machine learning algorithms typically used to map burned area is that they allow known sources of variation to be explicitly quantified. This makes statistical models well suited for wildfire detection with AVHRR data as the archive’s sources of noise are well documented. As such, we developed a methodology which uses autoregressive timeseries analysis to map circumboreal burned area with AVHRR data (Figure 1). Figure 1: Annual circumboreal wildfire extent mapped using data from the NOAA POES and ESA Metop-B AVHRR archive (1983-2020).
High wildfire years within the circumboreal region are usually facilitated by unseasonably hot and dry summertime conditions. Climate change is predicted to increase the severity of wildfire weather within the circumboreal region, which in turn is expected to increase annual burned area. However, it is unclear which circumboreal ecoregions are currently experiencing worsening wildfire weather, and to what degree this is affecting burned area extent within those ecoregions. We tested for long term trends in ecoregion level severe wildfire weather by conducting a timeseries analysis on circumboreal annual (1983-2020) growing season (Mar.-Nov.) 95th percentile Canadian Fire Weather Index value (Figure 2). We performed this timeseries analysis using a methodology known as remotePARTS which accounts for spatial and temporal autocorrelation in remotely sensed datasets. We then determined the impact of trends in severe wildfire weather on annual ecoregion level burned using autoregressive timeseries analysis. Figure 2: Long term trends in circumboreal 95th percentile growing season (March – November) Fire Weather Index value (1983-2020). Severe wildfire weather is worsening throughout the Eurasian boreal region, but shows greater spatial heterogeneity within the North American boreal region.
We found no significant long-term trends in severe wildfire weather within North American boreal ecoregions. However, wildfire weather appears to be worsening more rapidly in northern latitudes at the continental scale. Conversely, wildfire weather is worsening across all Eurasian boreal ecoregions with the Trans-Baikal conifer forests experiencing the greatest change. Trends in severe wildfire weather were also shown to influence interannual variability in burned area with the Northern Canadian Shield Taiga (50%) and Trans-Baikal Conifer Forests (30%) most sensitive to this relationship (Figures 3, 4). Figure 3: Percentage of interannual variability in North American boreal forest annual (1983-2020) growing season (March-November) burned area accounted for by growing season 95th percentile Fire Weather Index Value. Figure 4: Percentage of interannual variability in Eurasian boreal forest annual (1983-2020) growing season (March-November) burned area accounted for by growing season 95th percentile Fire Weather Index Value.
Climate change is causing the circumboreal wildfire season to become longer, drier, and warmer. Our findings highlight the ecoregion level differences in the intensity of this change. While no significant ecoregion level trends in burned area have yet been detected, the sensitivity of circumboreal wildfire regimes to wildfire weather makes them vulnerable to changing fire seasons. In particular, fire regimes of Eastern Eurasian boreal forests which are experiencing worsening wildfire weather are particularly vulnerable. Further study of climate and wildfire interaction within these ecoregions is required to understand how these fire regimes are currently changing, and how they may change in the future.
Classifying forests at tree species level from remotely sensed data over large areas is challenging, especially when ground-data do not exist. Since the opening of the Landsat archive in 2008, opportunities to improve forest type mapping and classification have increased, making it possible to explore phenological properties across different forest types. The seasonal dynamics of vegetation indices (e.g., enhanced vegetation index, EVI) are well correlated with the seasonal dynamics in photosynthetically active leaf area and are a good proxy for phenological stages. In addition to being helpful for individual forest type and tree species classification, phenology is linked to landscape resources because vegetation phenology determines food availability for a wide range of forest species.
To improve broad-scale forest mapping and landscape characterization, we developed an approach that can categorize forests based on both land surface phenology and climate characteristics, the forest phenoclusters. We calculated land surface phenology metrics based on EVI Sentinel-2 and EVI Landsat 8 combined annual time series. We also derived land surface temperature (LST) from Band 10 of the thermal infrared sensor (TIRS) of Landsat 8 and used precipitation from the WorldClim dataset. We then performed stratified X-means classification followed by hierarchical clustering. We applied the methodology in Argentina (2.8 million km2), which has a wide variety of forests, from rainforests to cold-temperate forests. We characterized the forest phenoclusters based on land surface phenology and climate characteristics, as well as based on strong regional expert knowledge.
We identified 54 forest phenoclusters across Argentina (Figure 1), each with unique combinations of vegetation phenology and climate characteristics. The resulting map is a valuable source of novel and ecologically relevant information applicable to management and conservation of biodiversity, for example, for stratifying biodiversity assessments, supporting wildlife habitat mapping and to improve landscape planning, including development of new reserves strategies. Additionally, using our method, it is possible to estimate phenoclusters at a variety of scales, which makes them useful for a variety of modeling applications. Specifically, forest phenoclusters could be an input data set to benefit species distribution modeling greatly over large areas with low data availability, such as for tapirs (Tapirus terrestris) and jaguars (Panthera onca) that occupy several ecoregions in Argentina. Figure 1. The 54 classes of forest phenoclusters across Argentina and some examples in five regions. ESP, Espinal; HDC, Humid and Dry Chaco; HM, High Monte; MAF, Misiones Atlantic; PAT, Continental Patagonian forests; PFS, Parana flooded savanna; SAY, Southern Andean Yungas; TDF, Patagonian forests of Tierra del Fuego.
Climate variability affects the phenology of vegetation and the seasonality of temperature, which can lead to mismatches between species and resources. When species are not able to track phenology and seasonal temperate changes, populations decline, posing a threat to biodiversity. Many plants and animals synchronize the timing of their life events with vegetation phenology. Mismatches in the timing of such events, can entail, for example, food limitations when the peak of bird nestling growth is not timed to occur during the annual peak in caterpillar abundance, which may affect reproductive success. In contrast, high spatial variability enhances ecological resilience to biodiversity loss from high inter-annual variability. Biodiversity should benefit from high spatial variability in vegetation greenness and land surface temperature within intact habitats, because such spatial variability is indicative of a variety of resources in close proximity and increases the likelihood that suitable conditions are available during times of extremes. Eduarda Silveira, a postdoctoral research in the SILVIS lab, recently published a study with her colleagues describing their efforts to identify hotspots of biodiversity conservation concern due to threats from high inter-annual variability (Figure 1). Figure 1. Potential integrations of inter-annual and spatial variability in vegetation greenness and land surface temperature, and the level of conservation concern for each integration.
They generated inter-annual and spatial remotely sensed indices based on time series analysis and image texture, respectively, and integrated these indices to identify areas of high, medium and low conservation concern (Figure 2). Figure 2. Areas of high and low conservation concern based on (1) vegetation greenness and (2) land surface temperature: (a) inter-annual variability in phenology, (b) spatial variability, (c) integration between inter-annual and spatial variability, and (d) hotspots maps.
They applied their method in Argentina. They identified hotspots of conservation concern in parts of northeastern and southern Argentina. These are sites where management efforts could be valuable (Figure 3). Eliminating existing pressures (i.e., dam construction, land use change) and improving spatial variability by increasing the abundance and diversity of natural landcover in these highly modified regions are promising approaches to increase resilience to climate extremes for native wildlife species. In contrast, areas in the northwest and central-west have high spatial variability, which may confer resilience to climate extremes, due to the variety of conditions and resources within close proximity (Figure 3). Adding protected areas in these naturally resilient regions may be effective in both protecting current patterns of biodiversity and maintaining their adaptive capacity to climate change. Eduarda Silveira hopes that her results will help Argentina’s conservation leaders to be strategic in their protection decision and to prioritize conservation management actions. Figure 3. Hotspots of highest and lowest conservation concern in Argentina.
Grassland ecosystems cover up to 40% of the global land area and provide many ecosystem services directly supporting the livelihoods of over 1 billion people. Monitoring long-term changes in grasslands is crucial for food security, biodiversity conservation, achieving Land Degradation Neutrality goals, and modeling the global carbon budget. Although long-term grassland monitoring using remote sensing is extensive, it is typically based on a single vegetation index and does not account for temporal and spatial autocorrelation, which means that some trends are falsely identified while others are missed. Our goal was to analyze trends in grasslands in Eurasia, the largest continuous grassland ecosystems on Earth. To do so, we calculated Cumulative Endmember Fractions (annual sums of monthly ground cover fractions) derived from MODIS 2002–2020 time series, and applied a new statistical approach PARTS that explicitly accounts for temporal and spatial autocorrelation in trends. We examined trends in green vegetation, non-photosynthetic vegetation, and soil ground cover fractions considering their independent change trajectories and relations among fractions over time. We derived temporally uncorrelated pixel-based trend maps and statistically tested whether observed trends could be explained by elevation, land cover, SPEI3, climate, country, and their combinations, all while accounting for spatial autocorrelation. We found no statistical evidence for a decrease in vegetation cover in grasslands in Eurasia. Instead, there was a significant map-level increase in non-photosynthetic vegetation across the region and local increases in green vegetation with a concomitant decrease in soil fraction. Independent environmental variables affected trends significantly, but effects varied by region. Overall, our analyses show in a statistically robust manner that Eurasian grasslands have changed considerably over the past two decades. Our approach enhances remote sensing-based monitoring of trends in grasslands so that underlying processes can be discerned.
