Cultivated land parcel recognition method based on deeplabv3+ semantic segmentation model for high-resolution unmanned aerial vehicle (UAV) imagery

The contemporary means of extracting arable land parcels based on semantic segmentation often neglects the fine-grained features of the parcels, which is poorly effective in dealing with irregularly shaped parcels, and is prone to the condition that the boundaries are not sufficiently clear and definite. To address the above problems, this study takes Xing'an County and Lipu County in Guilin City, Guangxi Zhuang Autonomous Region as the study area, and builds its own dataset of cropland plots, and proposes an improved deeplabv3+ semantic segmentation model for accurately extracting cultivated land parcels from mountainous UAV imagery, which employs MobileNetV2 as the backbone network, introduces the CBAM attention mechanism and the DFF dynamic feature fusion module, so that the extracted contour boundaries of cultivated plots more accurately. The improved deeplabv3+ model achieves 87.19% MPA and 80.36% MIoU in the extraction of cultivated plots. Comparison experiments with the deeplabv3+ and PSPNet models showed that the prediction accuracy of the improved model was better than that of the above two models, in which the MPA increased by 4.05% and 8.86%, and the MIoU increased by 5.62% and 10.17%. The extracted cropland plots are more consistent with the actual distribution pattern of cropland than the existing methods, which provides a practical method to accurately extract cropland information from irregular plots.