Automatically update land cover database from multi-temporal satellite images for ecological studies
Processing global and regional land cover time series changes are critical for climate, environmental and ecological studies, creating a pressing need to develop robust, efficient and accurate automated approaches for cost effective monitoring. The main challenges of using multi-temporal satellite data sources for updating land cover database changes include image inconsistency (in terms of spectral coverage, sensor calibration, and atmospheric correction), plant phenologic differences, weather variations, and difficulties of incorporating natural heterogeneous conditions into automatic image processing. In this study, we have been developing an innovative tool to automatically produce annual land cover changes using Landsat data sources. An existing land cover map for baseline classification and a corresponding cloud-free Landsat image used as a reference are requirements for this tool. The cloud-free Landsat reference image is used to improve the homogeneity within each subclass of the original baseline classification map through purification (i.e., removing outliers) and segmentation (i.e., dividing one class into several subclasses). Subsequently, time series Landsat images are used to detect annual changed and unchanged areas. Each pixel that is identified as unchanged inherits the baseline classification. However, for each pixel in the changed area, its classification is determined based on two major steps. First, pixels with similar remotely sensed metrics (e.g., reflectance, Normalized Difference Vegetation Index etc.) and conditions (e.g., elevation, temperature, precipitation, aspect etc.) within the unchanged areas are identified. Second, the classification from the majority of these identified pixels is assigned to the changed pixel.
This tool is being applied to three geographically different areas with land cover changes caused by forest mortality, crop rotation, and urbanization: Southern California (U.S.A.), the Great Plains (U.S.A.), and Beijing Olympic area (China). Preliminary results show our approach has resulted in promising classification accuracy. This tool can minimize the influence of data inconsistencies due to sensor, phenology, weather, and natural conditions. The tool was developed with very few user inputs and can be applied to each separate scene, enabling the feasibility to automatically update land cover database for ecological studies at regional and global scale.