Effects of Background Composition on Tobacco-Plant Segmentation Accuracy and Cross-Scene Stability in UAV RGB Imagery of Karst Fields

Abstract

In karst mountainous tobacco fields, white plastic mulch, bare rock, and weeds are spatially interspersed and can affect semantic segmentation accuracy in UAV RGB imagery. This study constructed seven background-composition scenarios from UAV RGB orthomosaics acquired in Zhenfeng County, Guizhou Province, China, and evaluated DeepLabV3+, U-Net, U-Net++, and U-Net Former in terms of overall performance, scene-specific accuracy, and cross-scene stability. DeepLabV3+ achieved the highest validation mIoU (0.91) during hyperparameter selection, whereas U-Net performed best on the spatially independent test set, with an mIoU of 0.86, a tobacco-class IoU of 0.74, and an OA of 0.99. Under single-background conditions, bare rock yielded the highest IoU for all four models, while the relative effects of white plastic mulch and weeds were model-dependent. Composite-background effects were non-monotonic: adding weeds to bare-rock scenes reduced IoU by 16.89–28.98%, whereas the three-background scenario did not yield the lowest IoU. U-Net achieved the highest scene-wise IoU in all seven scenarios and the lowest cross-scene coefficient of variation (8.06%). This study provides a reference for sample optimization, model selection, and error diagnosis in complex mountainous tobacco plant identification.