Mulched Film Mapping based on Contrastive Learning and Multi-Resolution Satellite Imagery

The plastic mulched film has been widely used in the agricultural farmland among arid and semi-arid areas, as it can effectively maintain soil temperature and reduce soil moisture evaporation. Recent studies show that plastic mulched film in farmland can be harmful to the soil structure and will cause soil degradation if they have note be recycled in time. Therefore, how to effectively monitor the use of plastic mulched film has become a key issue in current research. In the study, a contrastive learning-based method is proposed for mapping mulched film with multi-resolution satellite imagery. The proposed model contains two parts, as the common feature extraction part and the self-feature extraction part. Moreover, a pixel-scale mulched film satellite imagery and label dataset is built by in-situ investigation and manual labeling. The performance of the proposed method is evaluated among study area in Jiuquan, covering an area of 100 $\mathbf{k m}^{\mathbf{2}}$. Result shows the proposed method can well extract mulched film within the Sentinel-2 imagery of the study area, with F1-score at 0.92. The proposed method can be useful for agricultural management and environment protection.