Fine-grained classification of Orah mandarin tree plots in UAV remote sensing images based on GACL-DeepLabV3+

• Designed a Channel-aware Lightweight Spatial Attention. • Designed a Gated Axial Spatial Attention. • Designed a new pyramid structure, the Gated Axial Spatial Pyramid. • Proposed a method to classify Orah mandarin tree plots in mixed-age planting pattern. • The method achieves 94.73 % mIoU, 97.30 % mPA, 97.29 % mPrecision, and 97.17 % OA. The mixed planting of Orah mandarin trees of various ages is prevalent in the planting production areas owing to their long growth cycle. This exacerbates the fragmentation of the planting plots as well as affects the fruit farmers’ precise cultivation and management of Orah mandarin. Although progress has been made in the application of remote sensing technology for citrus planting area segmentation, most studies have focused on the overall identification of these areas. Studies on the fine-grained classification and extraction of Orah mandarin tree plots with varying tree ages and growth states within the same spatiotemporal range are lacking. This study proposed a GACL-DeepLabV3+ model based on high-resolution UAV imagery, incorporating Channel-aware Lightweight Spatial Attention and a Gated Axial Spatial Pyramid derived from Gated Axial Spatial Attention to accurately segment Orah mandarin plots and classify them by tree age and growth status, addressing blurred plot boundaries, feature similarity with other crops, and overlaps in exposed areas of young trees. Compared with other advanced deep learning models, GACL-DeepLabV3+ demonstrated notable performance in the classification of Orah mandarin tree plots. With only 5.876 M of parameters, this model achieved excellent performance with a mean Intersection over Union (mIoU) of 94.73 %, a mean pixel accuracy (mPA) of 97.30 %, a mean precision of 97.29 %, and an overall accuracy of 97.17 %, comprehensively verifying the accurate classification and extraction ability of the model in complex scenarios.