A Web-Based Prototype System for Deforestation Detection on High-Resolution Remote Sensing Imagery With Deep Learning

Forests are the largest carbon storage and carbon sinks on land and are closely related to global climate change. Using remote sensing images to monitor global deforestation has become the mainstream method, mainly because of their long time-series observation period and low cost. Although researchers have proposed several excellent deep-learning models to improve deforestation monitoring accuracy, few studies pay attention to designing an online deforestation detection system with deep-learning methods for nonexpert users or researchers, due to the complicated development process or excessive cost investment. In this study, we aimed to develop an open-sourced and user-friendly online system for monitoring large-scale deforestation on high-resolution remote sensing imageries, named OpenForestMonitor. The OpenForestMonitor system was the first online deforestation detection system developed for high-resolution remote sensing imageries with deep learning methods. To validate the performance of the OpenForestMonitor system, we selected four large-scale study areas in China as the study areas, and they are Liuyang City, Cangwu County, Sanya City, and Xianyang City. The final experiment results indicated that the F1-score of these study areas is 0.719, 0.854, 0.792, and 0.650, respectively. Furthermore, we also tested the performance of the OpenForestMonitor system on median-resolution Landsat remote sensing imageries and low-quality remote sensing images in complicated regions, and the final F1-score in Liuyang City and Cangwu County is 0.784 and 0.741, respectively. Compared with commercial software such as ArcGIS Pro 3.0, the OpenForestMonitor also indicated an obvious efficiency advantage. The OpenForestMonitor system was open-sourced, and it can also inspire researchers to develop other landcover, change detection, and object extraction online systems conveniently.