Maize Straw Identification Method for Conservation Tillage in Black Soil Area Based on UAV Imagery and SAM Weak Supervision Learning

Abstract:

Straw mulching is an important measure for conservation tillage in the black soil region. Straw identification is of great significance for the assessment of conservation tillage implementation effects and agricultural management decisions. To address the issue that fully supervised deep learning straw identification methods due to their reliance on a large amount of pixel-level labeled data, a weakly supervised learning straw identification method was proposed based on unmanned aerial vehicle (UAV) images and the segment anything model (SAM). By fine-tuned SAM with an adapter and a boundary-aware joint loss function, and generating high-quality pseudo-labels from bounding box weak annotations, an improved U-Net segmentation network was ultimately trained to achieve straw identification. A straw extraction experiment was conducted in the conservation tillage area of corn in Lishu County, Jilin Province as the research area. The results showed that the fine-tuned SAM achieved an MIoU of 81.04% and an F1-score of 87.85%, significantly outperforming the un-fine-tuned model. The model combining SAM weak supervision and the improved U-Net achieved a higher performance than other segmentation methods, with an F1-score of 90.6%. Ablation experiments verified the effectiveness of the joint loss function and convolutional modules in improving model performance. The research provided an efficient and cost-effective solution for remote sensing identification of straw identification in maize conservation tillage in the black soil region.