Multi-species weed density assessment based on semantic segmentation neural network.

The precise use of specific types of herbicides according to the weed species and density in fields can effectively reduce chemical contamination. A weed species and density evaluation method based on an image semantic segmentation neural network was proposed in this paper. A combination of pre-training and fine-tuning training methods was used to train the network. The pre-training data were images that only contain one species of weeds in one image. The weeds were automatically labeled by an image segmentation method based on the Excess Green (ExG) and the minimum error threshold. The fine-tuning dataset was real images containing multiple weeds and crops and manually labeled. Due to the limitation of computational resources, larger images were difficult to segment at one time. Therefore, this paper proposed a method of cutting images into sub-images. The relationship between sub-image size and segment accuracy was studied. The results showed that the training method reduced the workload of labeling training data while effectively avoiding overfitting. The accuracy of image segmentation decreases as the sub-image size decreases. Considering the limitation of computational resources, a subgraph of 256 × 256 size was selected. The semantic segmentation network achieved 97% overall accuracy. The coefficient of determination ( R 2 ) of weed density calculated by the algorithm and manually assessed was 0.90, and the root means square error (σ ) was 0.05. The method could effectively assess the density of each species of weeds in complex environments. It could provide a reference for accurate spraying herbicides based on weed density and species. [ABSTRACT FROM AUTHOR]