基于改进DeepLabV3+的栖霞市苹果园遥感识别OA

Remote sensing identification of apple orchards serves as a crucial foundation for the refined management of apple cultivation,yet it is prone to issues such as false detection,miseed detection,and blurred boundaries in complex land-cover contexts.To enhance identification accuracy,a high-resolution apple orchard dataset was constructed based on GF-2 imagery and field sampling data.An improved DeepLabV3+multi-level feature fusion model was proposed and applied to identify apple orchards in Qixia City.The model employs the lightweight MobileNetV2 as the backbone feature extraction network.By integrating the coordinate attention(CA)mechanism and strip pooling(SP)into the atrous spatial pyramid pooling(ASPP),a CASP-ASPP module was constructed to fuse multi-scale features.Additionally,an edge refinement module was introduced during the decoding stage to optimize boundary identification.Experimental results indicate that the improved model achieves an increase of 1.9 percentage points in mean intersection over union compared to the original model,and its overall identification accuracy outperformes various mainstream deep learning networks.This method can effectively improve the accuracy of remote sensing identification of apple orchards,providing reliable technical support for orchard monitoring and refined agricultural management.