Hongyan Liu∗,∗∗ Shun Ren∗,∗∗ Dong Ren∗,∗∗ and Xuan Liu∗,∗∗
Orchards segmentation, deep learning, remote sensing image, multi-scale category attention mechanism
Using remote sensing images to extract orchard is one of the important methods for product evaluation. Aiming at the problem of misclassification and omission caused by the semantic complexity of the marginal area and uneven planting density, a method based on the improved DeepLabv3 network model to segment the orchards is proposed. First, down-sampling is performed through the convolutional layer and residual module in the encoding network. Then, dilated convolution is used to build multi-scale modules, and category attention mechanism modules are added to refine high-level semantic features. Finally, the resolution of the feature map is restored using a deconvolution operation in the decoding network. Comparing the proposed method with the classic semantic segmentation network DeepLabv3 and other models on the validation set, the results show that the proposed method has better segmentation performance and generalization ability, and the segmentation result of the orchards is more accurate. To be accurate, the Mean Intersection over Union reaches 0.96, which is 2.3% more accurate than the original DeepLabv3+. The proposed method can more accurately segment the orchards in the unbalanced data set and provide a reference for product evaluation.
Important Links:
Go Back
