Parcel-Level Mapping of Artificial Forests Along the Middle Reach Valley of Yarlung Tsangpo River Based on Deep Learning Algorithms

Abstract

Artificial forest (AF) is an effective means of human intervention in forest ecosystems, aiming at preventing issues, such as soil erosion and land desertification. However, owing to the characteristics of large-scale afforestation projects, which often involve vast spatial extents and extended temporal scales, AF usually exhibits complex distribution patterns. In such cases, traditional remote sensing methods usually fail to accurately monitor AF conditions. To address this issue, this study introduced deep learning (DL) algorithms to extract multilevel features from remote sensing images for AF mapping and employed image processing techniques to enhance AF boundary determination. Through integrating these two approaches, high-resolution mapping of AF parcels was generated for a typical region in the middle reach valley of the Yarlung Tsangpo River. In the validation phase, the extracted regions were compared with manually labeled datasets and three accuracy metrics were calculated to demonstrate the extraction performance of the model. The accuracy reached 90.12% with the intersection over union (IoU) of 88.42%, and the cross-entropy loss function is only 0.0218. Meanwhile, three sampling areas with different coverages were selected for comparison, and the extractions have better performance than the SAM model based on the comparison with the samples. The findings reveal that this method can segment each AF parcel into independent objects, and the results would be helpful for parcel-based researches.