A Weakly-Supervised Learning Approach for RGB Crop Detection Using UAV Imagery

Precision agriculture increasingly relies on accurate crop monitoring to optimize yields and resource use, yet faces significant challenges due to the scarcity of labeled data and the high cost of advanced imaging technologies. To address these limitations, we propose a novel weakly-supervised learning approach for crop detection in aerial RGB imagery, in which pseudo-labels are automatically generated through zero-shot segmentation—thus minimizing the need for manual annotation. Our pipeline combines the Segment Anything Model for zero-shot segmentation, DBSCAN for clustering and label inference, and Faster R-CNN with a ResNet-101 backbone for object detection. This strategy enables effective crop detection even in scenarios with little or no human supervision, offering a scalable and cost-efficient solution for diverse agricultural contexts. We evaluate our approach using a newly collected dataset of drone-based RGB images, which comprises vineyards, orchards, olive groves, and wheat fields. Experimental results demonstrate high precision, recall, and F1 scores across crop types, validating the robustness and applicability of the proposed method in real-world agricultural environments. Our findings highlight the potential of integrating modern self-supervision techniques with object detection frameworks to enhance sustainable and data-efficient precision farming.