MSU-Net: Multi-Scale self-attention semantic segmentation method for oil-tea camellia planting area extraction in hilly areas of southern China

Abstract

Oil-tea camellia, one of the world’s four major edible woody oil trees, is acclaimed as the ’Oriental Olive Oil’ due to its exceptionally high nutritional value. The climate in southern China synchronizes with the ideal conditions for cultivating oil tea, making it the most abundant region globally in terms of its distribution. Consequently, the delineation of oil tea cultivation zones holds paramount significance for agricultural authorities in devising strategic production plans and management. However, the region is often affected by changeable weather and frequent cloud and rain, and there is a lack of continuous optical image data. Moreover, the complex topography primarily characterized by mountainous terrain, extensive coverage of farmlands, and vegetation has fragmented topographic features, posing challenges in accurately extracting semantic information from remote sensing images. To address these challenges, we propose a multi-scale self-attention semantic segmentation network aimed at meticulously identifying the semantic features of oil tea. Specifically, we introduce a self-attention mechanism to enable the model to comprehensively understand the information on feature images, followed by the integration of multi-scale feature images through the ASPP(Atrous Spatial Pyramid Pooloing) module to prevent the oversight of minor terrain features. Finally, the Dice-Loss function is applied to optimize the model’s segmentation of edge details. Experimental evaluations demonstrate that the proposed multi-scale self-attention semantic segmentation model achieved an Intersection over Union (IOU) of 0.93, Pixel Accuracy (PA) of 0.98, and Overall Accuracy (OA) of 94.83% for oil tea extraction on the dataset, showcasing a notable improvement over the original model. Additionally, we explore the method’s data requirements from the perspective of data volume and proportion. Ultimately, the experimental results demonstrate that our proposed method can accurately extract the oil tea cultivation areas in the cloudy and rainy hilly regions of southern China with high precision, thereby serving as a technological means for agricultural departments to oversee oil tea cultivation.