Efficient detection of corn straw coverage in complex agricultural scenarios

Abstract

Straw coverage serves as a critical indicator in the realm of conservation tillage. This study aims to fulfill the detection needs for straw coverage on edge monitoring platforms by initially capturing straw images through an onboard terminal and subsequently creating a dataset via data augmentation. We opted for SegNext as the foundational model and incorporated ResNet101 as the backbone to enhance the extraction of features specific to straw. To achieve a lightweight model without sacrificing detection accuracy, ResNet101 was utilized as the teacher model to mentor ResNet18 as the student model, with the training outcomes quantified using QAT. In tests conducted under multifactorial field scenarios, the QSR101-18 model achieved mIoU of 85.78 %, mAP of 95.98 % and Kappa of 86.25 %, surpassing SegNext by 1.44 %, 1.57 % and 1.32 %, respectively. The QSR101-18 model FLOPs and Params are 0.71G and 0.45 M respectively, which is about 1/27 and 1/100 of SegNext. When deployed on edge platforms and analyzed across varying straw coverage rates, QSR101-18 demonstrated an overall error of only 1.3 %, well within acceptable limits. The inference speed for a single image was just 16.32 ms, meeting the speed requirements for field operations. Consequently, the proposed QSR101-18 model demonstrates several key advantages, including a lightweight architecture, minimal error rates, robustness, and high accuracy. It effectively addresses the challenges posed by unstructured, fragmented straw and various environmental factors in detecting straw coverage, all while adhering to the speed constraints required for field operations on edge monitoring platforms.