Optimization of desert lake information extraction from remote sensing images using cellular automata

Abstract

Desert lakes are important wetland resources in the blown-sand area of western China and play a significant role in maintaining the regional ecological environment. However, large-scale coal mining in recent years has considerably impacted the deposition condition of several lakes. Rapid and accurate extraction of lake information based on satellite images is crucial for developing protective measures against desertification. However, the spatial resolution of these images often leads to mixed pixels near water boundaries, affecting extraction precision. Traditional pixel unmixing methods mainly obtain water coverage information in a mixed pixel, making it difficult to accurately describe the spatial distribution. In this paper, the cellular automata (CA) model was adopted in order to realize lake information extraction at a sub-pixel level. A mining area in Shenmu City, Shaanxi Province, China is selected as the research region, using the image of Sentinel-2 as the data source and the high spatial resolution UAV image as the reference. First, water coverage of mixed pixels in the Sentinel-2 image was calculated with the dimidiate pixel model and the fully constrained least squares (FCLS) method. Second, the mixed pixels were subdivided to form the cellular space at a sub-pixel level and the transition rules are constructed based on the water coverage information and spatial correlation. Lastly, the process was implemented using Python and IDL, with the ArcGIS and ENVI software being used for validation. The experiments show that the CA model can improve the sub-pixel positioning accuracy for lake bodies in mixed pixel image and improve classification accuracy. The FCLS-CA model has a higher accuracy and is able to identify most water bodies in the study area, and is therefore suitable for desert lake monitoring in mining areas.