Automatic Pear Extraction from High-Resolution Images by a Visual Attention Mechanism Network

Abstract

At present, forest and fruit resource surveys are mainly based on ground surveys, and the information technology of the characteristic forest and fruit industries is evidently lagging. The automatic extraction of fruit tree information from massive remote sensing data is critical for the healthy development of the forest and fruit industries. However, the complex spatial information and weak spectral information contained in high-resolution images make it difficult to classify fruit trees. In recent years, fully convolutional neural networks (FCNs) have been shown to perform well in the semantic segmentation of remote sensing images because of their end-to-end network structures. In this paper, an end-to-end network model, Multi-Unet, was constructed. As an improved version of the U-Net network structure, this structure adopted multiscale convolution kernels to learn spatial semantic information under different receptive fields. In addition, the “spatial-channel” attention guidance module was introduced to fuse low-level and high-level features to reduce unnecessary semantic features and refine the classification results. The proposed model was tested in a characteristic high-resolution pear tree dataset constructed through field annotation work. The results show that Multi-Unet was the best performer among all models, with classification accuracy, recall, F1, and kappa coefficient of 88.95%, 89.57%, 89.26%, and 88.74%, respectively. This study provides important practical significance for the sustainable development of the characteristic forest fruit industry.