Mapping Tea Plantations from Medium-Resolution Remote Sensing Images Using Convolutional Neural Networks and Swin Transformer

Tea is one of the world’s top three beverages, and the tea industry is an important pillar of China’s agricultural economy. The sustainable development of the tea industry requires rapid and accurate tea plantations mapping. In this paper, a novel deep learning model (TeaNet) was proposed to extract tea plantations from medium-resolution remote sensing images. The TeaNet model, which was designed as a U-shaped network structure, improved performance by coupled Swin Transformer and convolutional neural network (CNN). Furthermore, the Sentinel 2A images of Wuyishan City were utilized to validate the proposed model. The results indicated that the TeaNet model shows a good performance, with a recall of 82.75%, and an F1 score of 79.48% which outperforms the UNet (improved of 26.58% and 14.99%). This indicates that the TeaNet model can significantly overcome the interference of irrelevant information and reduce the edge adhesion of tea plantations, thereby identifying the planting areas of tea plantations and providing an effective method for large-scale tea plantation mapping.