基于优化HRNetV2的高分辨率遥感影像土地利用自动分类

研究目的：基于语义分割模型HRNetV2实现高分辨率遥感影像的土地利用自动分类，以期推动深度学习语义分割方法实现遥感影像的土地利用自动分类研究，并为该类研究提供可参考的案例。研究方法：首先对语义分割模型HRNetV2从损失函数、特征提取方面进行优化，以提高模型的分割精度；其次用优化后的HRNetV2模型对民乐县2 m的高分辨率遥感影像进行土地利用分类，并基于混淆矩阵对分类的结果进行评估。研究结果：（1）优化后的HRNetV2语义分割模型的MIOU达81.9%，相较于优化前提高了4.4%, 证明了优化方法的有效性；（2）针对民乐县进行的土地利用自动分类总体精度达89.72%，Kappa系数达0.888。研究结论：优化后的HRNetV2语义分割模型在高分辨率遥感影像的土地利用自动分类中具有较强的应用前景。

Automatic Land Use Classification Based on Optimized HRNetV2 High-resolution Remote Sensing Images

The purposes of this study are to promote the research of automatic land use classification of remote sensing images based on semantic segmentation model HRNetV2, and provide reference cases. The research methods are as follows. Firstly, the HRNetV2 semantic segmentation model was optimized from loss function and feature extraction to improve the segmentation accuracy. Then, the optimized HRNetV2 model was used to classify the 2-meter high-resolution remote sensing images of Minle County, and the classification results were evaluated based on confusion matrix. The results are as follows: 1) The MIOU of the optimized HRNetV2 semantic segmentation model is up to 81.9%, increasing by 4.4%, which proves the effectiveness of the optimized method. 2) The overall accuracy of automatic land use classification for Minle County is 89.72%, and the Kappa coefficient is 0.888. In conclusion, the optimized HRNetV2 semantic segmentation model has a strong application prospect in automatic land use classification of high-resolution remote sensing images.