A NEW WATER DETECTION FOR MULTISPECTRAL IMAGES BASED ON DATA SIMULATION AND RANDOM FOREST

In order to map surface water accurately and timely for different multispectral images, adequate and high-quality training samples are particularly important. However, the reusability of samples between different multispectral satellites and across geographic distance and time on classifying water is underdeveloped, which indicates that water bodies recognition of different images needs to build the corresponding water sample database. This study we addressed the limitation by using hyperspectral images to obtain high-quality training samples that could be used for water detection on different multispectral sensors. Combination of 12 spectral variables were used to classify surface water bodies using Random Forest classifier. The results showed that the overall accuracies were 96.12% for Sentinel-2 and 94.76% for Landsat 8, respectively. To compare the results of Sentinel-2 and Landsat 8 images had an overall similarity greater than 96%, these comparisons demonstrated that this algorithm had robustness to different multispectral sensors.