Weed Density Detection Method Based on a High Weed Pressure Dataset and Improved PSP Net

Abstract

Large-scale spraying on farmland is one of the most widely used weeding methods. Accurate weed density detection is of great significance for improving pesticide utilization and reducing environmental pollution. The purpose of this paper is to combine traditional image processing technology with deep learning technology to study semi-supervised annotation of the weed dataset and weed density detection method in a high-stress weed environment and provide a prescription map to guide variable spraying weeding operation. First, this paper uses a crop dataset to train U-Net to achieve crop segmentation and uses the color index and Otsu threshold segmentation algorithm to achieve vegetation segmentation. Then, weed segmentation is achieved by removing crop areas from vegetation segmentation results, and the segmentation results are made into a weed dataset. The improved PSP Net is trained using this dataset and weed segmentation is performed. The ratio of the number of weed pixels to the total number of pixels in the region is calculated for the obtained segmented image by region to measure the weed density. Finally, prescription maps representing different treatment intensities were generated based on the weed density threshold. Results indicate that EXG color index outperforms the other three indices for weed annotation. Compared with the original model, the MIoU, mPA, and Accuracy of the improved PSP Net model are increased by 2.15%, 0.92%, and 1.16%, respectively, and the model reasoning speed is increased by 6.9 times. The coefficient of determination between the predicted results of weed density and the manually labeled true values is 0.83, with a root mean square error of 0.17. The accuracy of the prescription map is 78%. The method proposed in this paper can effectively detect weed density in high-pressure weed environments and provide accurate prescription maps for variable spraying weed control.