Locally adaptive unmixing method for lake-water area extraction based on MODIS 250 m bands

Lakes in semi-arid and arid regions serve an important function in maintaining regional ecological balance. Thus, the change in lake-water area should be monitored by using remotely sensed images. However, most high-spatial-resolution satellite sensors cannot provide frequent observation data because of the long revisiting cycle and cloud effects. The 250 m MODIS images, as well as red and near-infrared bands, are currently the best monitoring data because of their frequent revisit and medium spatial resolution. Spectral unmixing is commonly used to extract information from coarse-resolution images at the subpixel level. However, in conventional unmixing, endmember selection is problematic, and a sufficient number of bands are necessary to solve the decomposition equations. In this study, we developed a locally adaptive unmixing (LAU) method to extract lake-water area using 250 m MODIS images. In this method, pixels mixed with water and land are initially extracted. Then, two classes of endmembers of each mixed pixel are determined by referring to the reflectivity of the neighboring pixels with different weights. Water abundance in each mixed pixel could be calculated using a single-band image. Owing to the overestimation in the NIR band and the underestimation in the red band, the average of the results from the two bands was set as the ultimate lake-water area to minimize error. This method is not only locally adaptive for endmember selection, but is also independent of the number of bands. This approach would be useful for the frequent monitoring of lake area by using 250 m MODIS images.