MSMRUnet: A Novel Framework for Mulched Film Mapping Based on Satellite Imagery and Deep Learning

Accurate information of the mulched film usage is an important mission within the remote sensing and agricultural society. However, due to the mixed pixel issue and noise from the satellite imagery, generating mulched film map for satellite imagery still remain challenging. In this work, a novel deep learning framework MSMRUnet is proposed for mapping mulched film from multi-source satellite imageries. The framework is designed to extract the multi-scale feature from both high-resolution satellite imagery and middle resolution satellite imagery. Specifically, the 0.8-meter spatial resolution BJ-2 satellite imagery and 20-meter spatial resolution Sentinel-2 satellite imagery of Wuwei City, Gansu Province are collected and used in this study. Result shows that the proposed MSMRUnet outperforms the comparison methods including traditional machine learning methods and common deep learning methods, with F1-score at 95.36. The proposed framework can be useful for agricultural plastic management, and farmland environment protection.