GREENHOUSE EXTRACTION FROM HIGH-RESOLUTION REMOTE SENSING IMAGERY WITH IMPROVED RANDOM FOREST

The timely and accurate acquisition of greenhouses and their distribution from remote sensing imagery is valuable for Chinese authorities seeking to optimize regional agricultural management and mitigate environmental pollution. However, greenhouses are uncommon background objects in such imagery, making them a minority class that traditional random forest (RF) methods struggle to classify accurately in unbalanced data sets. Herein, we propose and test an improved RF sample selection method. Equal sample numbers were randomly selected from minority and majority classes to build an original training set for RF modeling. High-quality samples were then automatically added to the training set according to the voting entropy and generalized Euclidean distance, which are based on sample characteristic parameters. The results demonstrate that our improved RF yields better results in identifying greenhouses than the traditional RF. In addition, our method can be utilized to identify other minority-class objects from remote sensing imagery.