Semantic segmentation of paddy image by UAV based on deep learning

In order to efficiently acquire water field information to improve precision agriculture applications, a semantic segmentation method of UAV water field images based on deep learning was proposed. Firstly, a set of high-resolution water field datasets were collected, and a bilateral filter was used to remove the image noise. Then, the encoder was adjusted to obtain more detailed field boundary feature information. Finally, the decoder was improved to fuse more shallow features and decouple the image depth information and spatial information by using depth separable convolution to obtain the DeepLabv3+ model with improved network structure. The experimental results showed that the pixel accuracy and average cross-merge ratio of the improved model were 96.04% and 85.90%, respectively, which were 2.09% and 4.66% better than the original model. Compared with the typical UNet, SegNet, and PSPNet semantic segmentation models, all the indexes had varying degrees of improvement. The method in this paper achieved accurate and efficient paddy field segmentation, which provided an important basis for obtaining paddy field boundary localization information and constructing high-precision farmland maps in the further.