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.
Over the course of a year, vegetation and temperature have strong phenological and seasonal patterns, respectively, and many species have adapted to these patterns. High inter-annual variability in the phenology of vegetation and in the seasonality of temperature pose a threat for biodiversity. However, areas with high spatial variability likely have higher ecological resilience where inter-annual variability is high, because spatial variability indicates presence of a range of resources, microclimatic refugia, and habitat conditions. The integration of inter-annual and spatial variability is thus important for biodiversity conservation. Areas where spatial variability is low and inter-annual variability is high are likely to limit resilience to disturbance. In contrast, areas of high spatial variability may be high priority candidates for protection. Our goal was to develop spatiotemporal remotely sensed indices to identify hotspots of biodiversity conservation concern. We generated indices that capture the inter-annual and spatial variability of vegetation greenness and land surface temperature and integrated them to identify areas of high, medium, and low biodiversity conservation concern. We applied our method in Argentina (2.8 million km2), a country with a wide range of climates and biomes. To generate the inter-annual variability indices, we analyzed MODIS Enhanced Vegetation Index (EVI) and Land Surface Temperature (LST) time series from 2001 to 2018, fitted curves to obtain annual phenological and seasonal metrics, and calculated their inter-annual variability. To generate the spatial variability indices, we calculated standard deviation image texture of Landsat 8 EVI and LST. When we integrated our inter-annual and spatial variability indices, areas in the northeast and parts of southern Argentina were the hotspots of highest conservation concern. High inter-annual variability poses a threat in these areas, because spatial variability is low. These are areas where management efforts could be valuable. In contrast, areas in the northwest and central-west are where protection should be strongly considered because the high spatial variability may confer resilience to disturbance, due to the variety of conditions and resources within close proximity. We developed remotely sensed indices to identify hotspots of high and low conservation concern at scales relevant to biodiversity conservation, = which can be used to target management actions in order to minimize biodiversity loss.
Environmental heterogeneity enhances species richness by creating niches and providing refugia. Spatial variation in climate has a particularly strong positive correlation with richness, but is often indirectly inferred from proxy variables, such as elevation and related topographic heterogeneity indices, or derived from interpolated coarsegrain weather station data. Our aim was to develop new remotely sensed metrics of relative temperature and thermal heterogeneity, compare them with proxy measures, and evaluate their performance in predicting species richness patterns. We analyzed Landsat 8’s Thermal Infrared Sensor data, calculated two thermal metrics during summer and winter, and compared their seasonal spatial patterns with those of elevation and topographic heterogeneity. We fit generalized least squares models to evaluate each variable’s effect in predicting seasonal bird richness using data from the North American Breeding Bird Survey. Generally speaking, neither elevation nor topographic heterogeneity were good proxies for temperature or thermal heterogeneity, respectively. Relative temperature had a non-linear relationship with elevation that was negatively quadratic in summer, but slightly positively quadratic in winter. Topographic heterogeneity had a stronger positive relationship with thermal heterogeneity in winter than in summer. The magnitude and direction of elevation–temperature and topographic heterogeneity–thermal heterogeneity relationships in each season also varied substantially across ecoregions. Remotely sensed metrics of relative temperature and thermal heterogeneity improved the predictive performance of species richness models, and both thermal variables had significant effects on bird richness that were independent of elevation and topographic heterogeneity. Thermal heterogeneity was positively related to total breeding bird richness, migrant breeding bird richness and resident bird richness, whereas topographic heterogeneity was negatively related to total breeding richness and unrelated to migrant or resident bird richness. Because thermal and topographic heterogeneity had contrasting seasonal patterns and effects on richness, they must be carefully contextualized when guiding conservation priorities.
After 1991, major events, such as the collapse of socialism and the transition to market economies, caused land use change across the former USSR and affected forests in particular. However, major land use changes may have occurred already during Soviet rule, but those are largely unknown and difficult to map for large areas because 30-m Landsat data is not available prior to the 1980s. Our goal was to analyze the rates and determinants of forest cover change from 1967 to 2015 along the Latvian-Russian border, and to develop an object-based image analysis approach to compare forest cover based on declassified Corona spy satellite images from 1967 with that derived from Landsat 5 TM and Landsat 8 OLI images from 1989/1990 and 2014/2015. We applied Structurefrom- Motion photogrammetry to orthorectify and mosaic the scanned Corona images, and extracted forest cover from Corona and Landsat mosaics using object-based image analysis in eCognition and expert classification. In a sensitivity analysis, we tested how the scale parameters for the segmentation affected the accuracy of the change maps. We analyzed forest cover and forest patterns for our full study area of 22,209 km2, and applied propensity score matching approach to identify three Latvian-Russian pairs of 15 × 15 km cells, which we compared. We attained overall classification accuracies of 92% (Latvia) and 93% (Russia) for the forest/non-forest change maps of 1967–1989, and 91% (Latvia) and 93% (Russia) for 1989–2015, and our results were robust in regards to the segmentation scale parameter. Sample-based forest cover gain from 1967 to 1989 differed notably between the two countries (18.5% in Latvia and 23.6% in Russia), but was generally much higher prior to 1989 than from 1989 to 2015 (8.7% in Latvia and 9.7% in Russia). Furthermore, we found rapid de-fragmentation of forest cover, where forest core area increased, and proportions of isolated patches and forest corridors decreased, and this was particularly pronounced in Russia. Our findings highlight the need to study Soviet-time land cover and land use change, because rural population declines and major policy decisions such as the collectivization of agricultural production, merging of farmlands and agricultural mechanization led already during Soviet rule to widespread abandonment and afforestation of remote farmlands. After 1991, government subsidies for farming declined rapidly in both countries, but in Latvia, new financial aid from the EU became available after 2001. In contrast, remoteness, lower population density, and less of a legacy of intensive cultivation resulted in higher rates of forest gain in Russia. Including Corona imagery in our object-based image analysis workflow allowed us to examine half a century of forest cover changes, and that resulted in surprising findings, most notably that forest area gains on abandoned farm fields were already widespread during the Soviet era and not just a postsocialist land use change trend as had been previously reported.
European Russia rapidly transitioned after the collapse of the Soviet Union from state socialism to a market economy. How did this political and economic transformation interact with ecological conditions to determine forest loss and gain? We explore this question with a study of European Russia in the two decades following the collapse of the Soviet Union. We identify three sets of potential determinants of forest-cover change—supply-side (environmental), demand-side (economic), and political/administrative factors. Using new satellite data for three distinct types of forest-cover change—logging, forest fires, and forest gain—we quantify the relative importance of these variables in province-level regression models during periods of a) state collapse (1990s), and b) state growth (2000s). The three sets of covariates jointly explain considerable variation in the outcomes we examine, with size of forest bureaucracy, autonomous status of the region, and prevalence of evergreen forests emerging as robust predictors of forest-cover change. Overall, economic and administrative variables are significantly associated with rates of logging and reforestation, while environmental variables have high explanatory power for patterns of forest fire loss.
Cropland abandonment is a widespread land-use change, but it is difficult to monitor with remote sensing because it is often spatially dispersed, easily confused with spectrally similar land-use classes such as grasslands and fallow fields, and because post-agricultural succession can take different forms in different biomes. Due to these difficulties, prior assessments of cropland abandonment have largely been limited in resolution, extent, or both. However, cropland abandonment has wide-reaching consequences for the environment, food production, and rural livelihoods, which is why new approaches to monitor long-term cropland abandonment in different biomes accurately are needed. Our goals were to 1) develop a new approach to map the extent and the timing of abandoned cropland using the entire Landsat time series, and 2) test this approach in 14 study regions across the globe that capture a wide range of environmental conditions as well as the three major causes of abandonment, i.e., social, economic, and environmental factors. Our approach was based on annual maps of active cropland and non-cropland areas using Landsat summary metrics for each year from 1987 to 2017. We streamlined perpixel classifications by generating multi-year training data that can be used for annual classification. Based on the annual classifications, we analyzed land-use trajectories of each pixel in order to distinguish abandoned cropland, stable cropland, non-cropland, and fallow fields. In most study regions, our new approach separated abandoned cropland accurately from stable cropland and other classes. The classification accuracy for abandonment was highest in regions with industrialized agriculture (area-adjusted F1 score for Mato Grosso in Brazil: 0.8; Volgograd in Russia: 0.6), and drylands (e.g., Shaanxi in China, Nebraska in the U.S.: 0.5) where fields were large or spectrally distinct from non-cropland. Abandonment of subsistence agriculture with small field sizes (e.g., Nepal: 0.1) or highly variable climate (e.g., Sardinia in Italy: 0.2) was not accurately mapped. Cropland abandonment occurred in all study regions but was especially prominent in developing countries and formerly socialist states. In summary, we present here an approach for monitoring cropland abandonment with Landsat imagery, which can be applied across diverse biomes and may thereby improve the understanding of the drivers and consequences of this important land-use change process.
The United States (U.S.) federal government provides imagery obtained by federally funded Earth Observation satellites typically at no cost. For many years Landsat was an exception to this trend, until 2008 when the United States Geological Survey (USGS) made Landsat data accessible via the internet for free. Substantial increases in downloads of Landsat imagery ensued and led to a rapid expansion of science and operational applications, serving government, private sector, and civil society. The Landsat program hence provides an example to space agencies worldwide on the value of open access for Earth Observation data and has spurred the adaption of similar policies globally, including the European Copernicus Program. Here, we describe important aspects of the Landsat free and open data policy and highlight the importance and continued relevance of this policy.
Maximum Arctic sea ice extent (million km^2) during the month of September (1979 – 2018).
Rapid warming of the Arctic is expected to cause substantial loss arctic sea ice volume and surface area. This will have numerous ecological and economic implications including habitat loss for sensitive species such as Polar Bear and Walrus, accelerated heating of the Arctic Ocean due to lowered albedo, and heightened shipping activity within the Arctic due to the increased availability of ice-free ports.
Sea ice loss has already been rapid, particularly during the summer months (Figure 1), and that this trend will only accelerate. The question is when and when exactly these losses occurred, and if observed trends are statistically significant. A major reason why this is difficult to ascertain is spatial and temporal autocorrelation, that is the influence nearby ice, and of previous ice abundance has on current conditions. We account for autocorrelation in our analyses of Arctic sea ice loss using a constrained least squares autoregressive model that predicts ice abundance based on the prior year’s conditions.
We found that three regions have experienced significant and highly temporally autocorrelated sea ice loss; a) the Laptev Sea near Severnaya Zemlya, Russia, b) the Beaufort Sea near northeastern Alaska, USA, and c) the Arctic Ocean near Prince Patrick island, Canada (Maps 1, 2, 3). Our next step is to apply a Generalized Least Squares model intended to detect spatial patterns in trends of sea ice loss.
T-Scores corresponding to trends in September Arctic sea ice extent (1979 – 2020) Negative values correspond to decreased sea ice abundance.Temporal autocorrelation in September Arctic sea ice extent (1979 – 2020). Extreme values correspond to higher degrees of temporal autocorrelation in trend of sea ice abundance.Mean days of sea ice presence for the month of September (1979 – 2020) Negative values correspond to decreased sea ice abundance.
Human activity cause major changes to the planet and biodiversity is declining at an alarming rates. In order to prevent biodiversity loss, conservation actions require to assess current status of biodiversity to better understand and predict future changes, to identify the major drivers of biodiversity patterns, and to map biodiversity patterns. However, monitoring biodiversity over large areas is challenging to do in the field. Remote sensing provides the opportunity to develop indices that are designed for biodiversity assessment, because satellite data are collected systematically across broad scales. Vegetation productivity is one of the important determinants of species richness and density across broad scale. Vegetation indices derived from satellite data are a good proxy for vegetation productivity over broad areas. The Dynamic Habitat Indices (DHIs) summarize the three different aspects of vegetation productivity: cumulative productivity, minimum productivity, and seasonality in the way that it became relevant for biodiversity (Hobi et al., 2017; Radeloff et al., 2019; Razenkova et al., 2020). However, so far the DHIs have only been derived from coarse-resolution satellite imagery, which limits their value for management decisions.
The DHIs based on Landsat 8 in Colorado US, spatial resolution 30-m.The DHIs based on MODIS in Colorado US, spatial resolution 1-km.
Our goal is to develop the DHIs using medium-resolution Landsat imagery for monitoring biodiversity and abundance pattern across the conterminous United States. The main advantage is that imagery with medium resolution provides more detailed information about the spatial patterns of productivity. Our rationale was that the DHIs with higher spatial resolution could capture the difference in vertical structure of vegetation and characterize habitat heterogeneity at much finer scale, especially in complex mountainous terrain and areas with fragmented land cover. Having this crucial information in my hand, will help to understand how species respond to anthropogenic modification of landscapes, which disturb the integrity of landscape pattern. However, the temporal resolution of Landsat is low, and that creates a lot of challenges for the calculation of the DHIs.
We will develop the DHIs for the conterminous United States and test the usefulness of the DHIs for explaining the avian species richness and abundance pattern. Our study covers a wide range of ecoregions, and has diverse climatic zones and topography, resulting in a large number of habitats and large ranges of the DHIs. Moreover, rich datasets for bird richness and abundance are available for the US, particularly the western US. Our research will add more understanding to the importance of higher spatial resolution for characterizing the DHIs metrics and consequently for modeling biodiversity and individual species pattern. Moreover, our work will add more knowledge about drivers of avian diversity across broad spatial extents that can be used to predict how biodiversity patterns will change in the future depending on changes in vegetation productivity.
Southern California wildfires captured from the International Space Station. Image Credit: NASA Johnson
The wildland-urban interface (WUI), the area where houses and wildland vegetation mix or intermingle, is rapidly growing. Housing expansion into the WUI has increased wildfire frequency and risk. This is in part because human-caused wildfires are responsible for most of the wildfires in the WUI. The rising intensity and frequency of wildfires have led to increased damage and loss of life in recent years. Because wildfire risk increases when people live near or in the WUI, it is beneficial to study historical housing trends and urbanization in these areas.
Globally, more frequent wildfires have become the new normal, and it is important to understand if and how housing development has worsened the damage. To find answers, Neda is using historic and current high-resolution satellite imagery to map housing units in California over the past 50 years. Neda plans to evaluate housing growth from the 1970s to the present-day to evaluate if there has been a proportional increase or decrease of houses in California in the WUI.
Anu Kramer (former SILVIS lab member) digitized total buildings destroyed by 89 wildfires in California from 1985-2014 and the Tubbs Fire in 2017. Figure from H.A. Kramer et al. (2019)
Neda’s main objective is to identify the areas of increased housing in the WUI and, ultimately, find out whether these areas have adapted to potential risks caused by wildfires. The main idea is to provide insight into how humans have adapted to wildfires over the past 50 years. Researching whether policy and communities adjust where houses are built based on wildfire risk could provide insight into the extent of WUI growth.
With thoughtful housing choices, there could be substantial benefits to the environment, as well as saving human lives, money, houses, and wild animals. Since it is always wise to “learn from our own mistakes,” it is important to understand which choices in the 20th century have caused additional wildfire risk. Stay tuned and follow Neda’s work for insight into thoughtful housing choices in California.
Examples of intermix (left) and interface (right) wildland-urban interface areas located near Redding, California. Photos from Google Earth.
Sarybash village in Azerbaijan and its surroundings with fragments of woody vegetation. Left image: Corona; Right image: courtesy Azerchin Muradov
Humans have altered natural landscapes for millennia, but especially so in recent decades. Recognizing how landscapes have changed over time is critical to understanding the relationship between humans and their environment. In studying this relationship, detailed land cover maps are a key resource. However, to date, there are few detailed land cover maps for the mid-20th century due to of the minimal amount of spatial data available during this time. Afag is developing a novel approach to detailed land cover mapping that will provide more context to mid-20th century landscape changes. Her final maps will incorporate land cover classes such as forest, grassland, agricultural land, water, and urban area.
Gold mining in Sakdrisi mountain slopes, Kazreti, Georgia. Left image: Corona; Right image: recent Google imagery
Afag is working on a methodology that will automatically map land cover maps anywhere on the globe from high resolution black and white imagery captured by Corona satellites, i.e., former US spy satellites. Afag is testing her methodology in the Caucasus eco-region encompassing the countries of Azerbaijan, Russia, Georgia, and Armenia, which has diverse ecosystem. This region has areas of extreme elevation and unique biomes. That means that if her methods work in this very diverse region, they should be applicable anywhere. Afag is using object-based image analysis to extract important land use information from satellite data. The classifications are conducted in Google Earth Engine and results in detailed land cover maps.
Growth of Aghdash city in Azerbaijan. Left image: Corona; Right image: recent Google imagery
Land cover maps are a valuable for both scientists and land managers. They provide means to evaluate how past land use decisions have affected the current land cover. The maps may also be helpful for projecting future land use changes. Additionally, land use maps can assist in identifying the scale at which politics and wars affect the landscape over time. Ultimately, Afag’s methodology will have the capability to automatically map mid-20th century land cover in large areas across the globe. Her research will help solve tough questions relating to landscape changes over the last half-century.
Altınçayır creek dam in Turkey constructed in 1995-2001. Left image: Corona; Right image: recent Google imagery
