Coastal forest cover change detection using satellite images and convolutional neural networks in Vietnam

Abstract

Monitoring forest cover changes is an important task for forest resource management and planning. In this context, remote sensing images have shown a high potential in forest cover changes detection. In Vietnam, although the existence of a large number of such images and ground-truth labels, current researches still relied on classical methods employed manual indices, such as multi-variant change vector analysis (MVCA) and normalized difference vegetation index. These methods highly require domain knowledge to determine threshold values for forest change that are applicable only for studied areas. Therefore, in this paper, we propose a method to detect coastal forest cover changes, which can exploit available dataset and ground-truth labels. Moreover, the proposed method does not require much domain knowledge. We used multi-temporal Sentinel-2 imagery to train a segmentation model, that is based on the U-Net network. It was used then to detect forest areas at the same location taken at different times. Lastly, we compared obtained results to identify forest disturbances. Experimental results demonstrated that our method provided a high accuracy of 95.4% on the testing set. Furthermore, we compared our model with the MVCA method and found that our model outperforms this popular method by 3.8%.