Detection of ground straw coverage under conservation tillage based on deep learning

Abstract

Given the technical requirement for the fast acquisition of ground straw coverage from conservation tillage fields, we investigated the accurate detection of straw coverage by using unmanned aerial vehicle (UAV) low-altitude remote sensing images. Ground images of farmlands under conservation tillage were acquired using low-altitude UAV. A semantic segmentation algorithm of straw coverage was established with the improved U-Net on ResNet18. The algorithm was evaluated and compared with VGG11–U-Net (the algorithm with VGG11 as the U-Net extraction module), triangle algorithm (TRIANGLER) and ‘RGB, HSV and grey + support vector machine (SVM)’ (the algorithm based on multicolour space information fusion combined with SVM). The mean absolute deviation of field straw coverage determined by the new algorithm was 3.56%, which was lower than that by other algorithms. Thus, the new algorithm offers data for the rapid acquisition of ground straw coverage from farmlands under conservation tillage.