CSO-YOLO: A Cross-Species Occlusion-Aware Object Detection Framework for Soybean–Maize Strip Intercropping Fields

Abstract

Soybean–maize strip intercropping improves landuse efficiency and crop productivity by exploiting complementary crop structures and resource use. Accurate detection of maize ears and soybean pods is important for growth monitoring, yield estimation, and intelligent agricultural machinery. However, natural field images contain cross-crop occlusion, complex backgrounds, and large differences in target scale and shape, which make joint detection challenging. To address these problems, this paper proposes CSO-YOLO, a cross-species occlusion-aware object detection framework. The framework enhances discriminative target-region representations under occlusion through a Cross-Species Occlusion Attention (CSOA) module, promotes the interaction between shallow spatial details and deep semantic information through bidirectional multi-scale feature fusion, and improves the localization of dense soybean pods through a highresolution small-object detection branch. Experimental results show that CSO-YOLO achieves 93.0% mAP@0.5 on the test set, with AP values of 97.5% and 88.6% for soybean pods and maize ears, respectively. These results indicate that CSO-YOLO can effectively detect maize ears and soybean pods in complex soybean–maize strip intercropping fields.