Patrick Hostert

Sensitivity of Support Vector Machines to Random Feature Selection in Classification of Hyperspectral Data

The accuracy of supervised land cover classifications depends on factors such as the chosen classification algorithm, adequate training data, the input data characteristics, and the selection of features. Hyperspectral imaging provides more detailed spectral and spatial information on the land cover than other remote sensing resources. Over the past ten years, traditional and formerly widely accepted statistical classification methods have been superseded by more recent machine learning algorithms, e.g., support vector machines (SVMs), or by multiple classifier systems (MCS). This can be explained by limitations of statistical approaches with regard to high-dimensional data, multimodal classes, and often limited availability of training data. In the presented study, MCSs based on SVM and random feature selection (RFS) are applied to explore the potential of a synergetic use of the two concepts. We investigated how the number of selected features and the size of the MCS influence classification accuracy using two hyperspectral data sets, from different environmental settings. In addition, experiments were conducted with a varying number of training samples. Accuracies are compared with regular SVM and random forests. Experimental results clearly demonstrate that the generation of an SVM-based classifier system with RFS significantly improves overall classification accuracy as well as producers and users accuracies. In addition, the ensemble strategy results in smoother, i.e., more realistic, classification maps than those from stand-alone SVM. Findings from the experiments were successfully transferred onto an additional hyperspectral data set.

The EnMAP Spaceborne Imaging Spectroscopy Mission for Earth Observation

Imaging spectroscopy, also known as hyperspectral remote sensing, is based on the characterization of Earth surface materials and processes through spectrally-resolved measurements of the light interacting with matter. The potential of imaging spectroscopy for Earth remote sensing has been demonstrated since the 1980s. However, most of the developments and applications in imaging spectroscopy have largely relied on airborne spectrometers, as the amount and quality of space-based imaging spectroscopy data remain relatively low to date. The upcoming Environmental Mapping and Analysis Program (EnMAP) German imaging spectroscopy mission is intended to fill this gap. An overview of the main characteristics and current status of the mission is provided in this contribution. The core payload of EnMAP consists of a dual-spectrometer instrument measuring in the optical spectral range between 420 and 2450 nm with a spectral sampling distance varying between 5 and 12 nm and a reference signal-to-noise ratio of 400:1 in the visible and near-infrared and 180:1 in the shortwave-infrared parts of the spectrum. EnMAP images will cover a 30 km-wide area in the across-track direction with a ground sampling distance of 30 m. An across-track tilted observation capability will enable a target revisit time of up to four days at the Equator and better at high latitudes. EnMAP will contribute to the development and exploitation of spaceborne imaging spectroscopy applications by making high-quality data freely available to scientific users worldwide.

Bringing an ecological view of change to Landsat-based remote sensing

When characterizing the processes that shape ecosystems, ecologists increasingly use the unique perspective offered by repeat observations of remotely sensed imagery. However, the concept of change embodied in much of the traditional remote-sensing literature was primarily limited to capturing large or extreme changes occurring in natural systems, omitting many more subtle processes of interest to ecologists. Recent technical advances have led to a fundamental shift toward an ecological view of change. Although this conceptual shift began with coarser-scale global imagery, it has now reached users of Landsat imagery, since these datasets have temporal and spatial characteristics appropriate to many ecological questions. We argue that this ecologically relevant perspective of change allows the novel characterization of important dynamic processes, including disturbances, longterm trends, cyclical functions, and feedbacks, and that these improvements are already facilitating our understanding of critical driving forces, such as climate change, ecological interactions, and economic pressures.

Pixel-Based Image Compositing for Large-Area Dense Time Series Applications and Science

Abstract Free and open access to the more than 40 years of data captured in the Landsat archive, combined with improvements in standardized image products and increasing computer processing and storage capabilities, have enabled the production of large-area, cloud-free, surface reflectance pixel-based image composites. Best-available-pixel (BAP) composites represent a new paradigm in remote sensing that is no longer reliant on scene-based analysis. A time series of these BAP image composites affords novel opportunities to generate information products characterizing land cover, land cover change, and forest structural attributes in a manner that is dynamic, transparent, systematic, repeatable, and spatially exhaustive. Herein, we articulate the information needs associated with forest ecosystem science and monitoring in a Canadian context, and indicate how these new image compositing approaches and subsequent derived products can enable us to address these needs. We highlight some of the issues and opportunities associated with an image compositing approach and demonstrate annual composite products at a national-scale for a single year, with more detailed analyses for two prototype areas using 15 years of Landsat data. Recommendations concerning how to best link compositing decisions to the desired use of the composite (and the information need) are presented, along with future research directions. Résumé L’accès libre et gratuit à plus de 40 ans de données dans l’archive Landsat combiné à l’amélioration des produits d’imagerie standardisés et l’augmentation des capacités de traitement et de stockage informatiques ont permis la production d’images composites basées sur les pixels de réflectance de surface de grande superficie sans nuages. Les composites du « meilleur pixel disponible » (best-available-pixel; BAP) représentent un nouveau paradigme en matière de télédétection qui ne dépend plus de l’analyse par scène. Une série chronologique de ces images composites BAP offre de nouvelles occasions de générer des produits d’information qui caractérisent la couverture terrestre, le changement de la couverture terrestre et les attributs structurels de la forêt d’une manière dynamique, transparente, systématique, répétable et spatialement exhaustive. Ici, nous articulons les besoins d’information liés à la science et à la surveillance des écosystèmes forestiers dans un contexte canadien, et nous indiquons comment ces nouvelles approches de composition d’image et les produits qui en découlent peuvent nous permettre de répondre à ces besoins. Nous soulignons quelques-uns des problèmes et des possibilités associés à une approche de composition d’image et nous démontrons des produits composites annuels à l’échelle nationale pour une année, avec des analyses plus détaillées pour deux zones prototypes utilisant 15 ans de données Landsat. Des recommandations concernant la meilleure façon de lier des décisions de composition d’images à l’utilisation souhaitée du composite (et le besoin d’information) ainsi que les orientations futures de la recherche sont présentées.

Coupling spectral unmixing and trend analysis for monitoring of long-term vegetation dynamics in Mediterranean rangelands

The development of vegetation cover is one of the primary indicators for land degradation, stability, or regeneration in regions threatened by overgrazing. This paper addresses the problem how spatially explicit information about degradation processes in European Mediterranean rangelands can be derived from long time series of satellite data. The selected test site in central Crete, Greece, is considered to be representative for the highly heterogeneous character of such landscapes. The monitoring approach comprises the time period between 1977 and 1996, covered by nine Landsat TM and four Landsat MSS images. Special emphasis has hence been put on the evaluation of potentials and drawbacks when coupling Landsat TM and MSS based results. The data sets were geometrically and radiometrically pre-processed in a rigorous fashion, followed by a linear spectral unmixing approach and a time series analysis of vegetation fraction images. Based on the resulting map, the spatio-temporal patterns of vegetation cover changes are explained. Even a test site such as central Crete, with its limited spatial extend, exhibits heterogeneous patterns of change, supporting the hypothesis that long time series of EOS data from Landsat-like sensors are mandatory to identify the relevant changes at landscape level.

POST‐SOCIALIST FOREST DISTURBANCE IN THE CARPATHIAN BORDER REGION OF POLAND, SLOVAKIA, AND UKRAINE

Forests provide important ecosystem services, and protected areas around the world are intended to reduce human disturbance on forests. The question is how forest cover is changing in different parts of the world, why some areas are more frequently disturbed, and if protected areas are effective in limiting anthropogenic forest disturbance. The Carpathians are Eastern Europes largest contiguous forest ecosystem and are a hotspot of biodiversity. Eastern Europe has undergone dramatic changes in political and socioeconomic structures since 1990, when socialistic state economies transitioned toward market economies. However, the effects of the political and economic transition on Carpathian forests remain largely unknown. Our goals were to compare post-socialist forest disturbance and to assess the effectiveness of protected areas in the border triangle of Poland, Slovakia, and Ukraine, to better understand the role of broadscale political and socioeconomic factors. Forest disturbances were assessed using the forest disturbance index derived from Landsat MSS/TM/ETM+ images from 1978 to 2000. Our results showed increased harvesting in all three countries (up to 1.8 times) in 1988-1994, right after the system change. Forest disturbance rates differed markedly among countries (disturbance rates in Ukraine were 4.5 times higher than in Poland, and those in Slovakia were 4.3 times higher than in Poland), and in Ukraine, harvests tended to occur at higher elevations. Forest fragmentation increased in all three countries but experienced a stronger increase in Slovakia and Ukraine (approximately 5% decrease in core forest) than in Poland. Protected areas were most effective in Poland and in Slovakia, where harvesting rates dropped markedly (by nearly an order of magnitude in Slovakia) after protected areas were designated. In Ukraine, harvesting rates inside and outside protected areas did not differ appreciably, and harvests were widespread immediately before the designation of protected areas. In summary, the socioeconomic changes in Eastern Europe that occurred since 1990 had strong effects on forest disturbance. Differences in disturbance rates among countries appear to be most closely related to broadscale socioeconomic conditions, forest management practices, forest policies, and the strength of institutions. We suggest that such factors may be equally important in other regions of the world.

A Pixel-Based Landsat Compositing Algorithm for Large Area Land Cover Mapping

Information on the changing land surface is required at high spatial resolutions as many processes cannot be resolved using coarse resolution data. Deriving such information over large areas for Landsat data, however, still faces numerous challenges. Image compositing offers great potential to circumvent such shortcomings. We here present a compositing algorithm that facilitates creating cloud free, seasonally and radiometrically consistent datasets from the Landsat archive. A parametric weighting scheme allows for flexibly utilizing different pixel characteristics for optimized compositing. We describe in detail the development of three parameter decision functions: acquisition year, day of year and distance to clouds. Our test site covers 42 Landsat footprints in Eastern Europe and we produced three annual composites. We evaluated seasonal and annual consistency and compared our composites to BRDF normalized MODIS reflectance products. Finally, we also evaluated how well the composites work for land cover mapping. Results prove that our algorithm allows for creating seasonally consistent large area composites. Radiometric correspondence to MODIS was high (up to R2 > 0.8), but varied with land cover configuration and selected image acquisition dates. Land cover mapping yielded promising results (overall accuracy 72%). Class delineations were regionally consistent with minimal effort for training data. Class specific accuracies increased considerably (~10%) when spectral metrics were incorporated. Our study highlights the value of compositing in general and for Landsat data in particular, allowing for regional to global LULCC mapping at high spatial resolutions.

Agricultural land change in the Carpathian ecoregion after the breakdown of socialism and expansion of the European Union

Widespread changes of agricultural land use occurred in Eastern Europe since the collapse of socialism and the European Union’s eastward expansion, but the rates and patterns of recent land changes remain unclear. Here we assess agricultural land change for the entire Carpathian ecoregion in Eastern Europe at 30 m spatial resolution with Landsat data and for two change periods, between 1985–2000 and 2000–2010. The early period is characterized by post-socialist transition processes, the late period by an increasing influence of EU politics in the region. For mapping and change detection, we use a machine learning approach (random forests) on image composites and variance metrics which were derived from the full decadal archive of Landsat imagery. Our results suggest that cropland abandonment was the most prevalent change process, but we also detected considerable areas of grassland conversion and forest expansion on non-forest land. Cropland abandonment was most extensive during the transition period and predominantly occurred in marginal areas with low suitability for agriculture. Conversely, we observed substantial recultivation of formerly abandoned cropland in high-value agricultural areas since 2000. Hence, market forces increasingly adjust socialist legacies of land expansive production and agricultural land use clusters in favorable areas while marginal lands revert to forest.

Mapping the extent of abandoned farmland in Central and Eastern Europe using MODIS time series satellite data

The demand for agricultural products continues to grow rapidly, but further agricultural expansion entails substantial environmental costs, making recultivating currently unused farmland an interesting alternative. The collapse of the Soviet Union in 1991 led to widespread abandonment of agricultural lands, but the extent and spatial patterns of abandonment are unclear. We quantified the extent of abandoned farmland, both croplands and pastures, across the region using MODIS NDVI satellite image time series from 2004 to 2006 and support vector machine classifications. Abandoned farmland was widespread, totaling 52.5 Mha, particularly in temperate European Russia (32 Mha), northern and western Ukraine, and Belarus. Differences in abandonment rates among countries were striking, suggesting that institutional and socio-economic factors were more important in determining the amount of abandonment than biophysical conditions. Indeed, much abandoned farmland occurred in areas without major constraints for agriculture. Our map provides a basis for assessing the potential of Central and Eastern Europe’s abandoned agricultural lands to contribute to food or bioenergy production, or carbon storage, as well as the environmental trade-offs and social constraints of recultivation.

Classifying segmented hyperspectral data from a heterogeneous urban environment using support vector machines

Classifying remotely sensed images from urban environments is challenging. Urban land cover classes are spectrally heterogeneous and materials from different classes have similar spectral properties. Image segmentation has become a common preprocessing step that helped to overcome such problems. However, little attention has been paid to impacts of segmentation on the datas spectral information content. Here, urban hyperspectral data is spectrally classified using support vector machines (SVM). By training a SVM on pixel information and applying it to the image before segmentation and after segmentation at different levels, the classification framework is maintained and the influence of the spectral generalization during image segmentation hence directly investigated. In addition, a straightforward multi-level approach was performed, which combines information from different levels into one final map. A stratified accuracy assessment by urban structure types is applied. The classification of the unsegmented data achieves an overall accuracy of 88.7%. Accuracy of the segment-based classification is lower and decreases with increasing segment size. Highest accuracies for the different urban structure types are achieved at varying segmentation levels. The accuracy of the multi-level approach is similar to that of unsegmented data but comprises the positive effects of more homogeneous segment-based classifications at different levels in one map.