ResNet和EfficientNet遥感图像场景分类研究

Study on ResNet and EfficientNet Remote Sensing Image Scene Classification

摘要: 针对遥感图像分类问题，首先采用了ResNet模型和EfficientNet模型进行训练，其中，前者采用18层网络结构，并在测试集得到了59.8%的准确度；后者采用了EfficientNet-B0和EfficientNet-B1模型，最高分类精度为92.6%。在此基础上，使用Ranger优化器替换了SGD方法，虽然测试集精度与原方法近似，但使用Ranger优化器训练EfficientNet模型具有更强的稳定性和更快的训练速度。最后，对测试的结果进行了分析，列举出每一类场景的分类精度，分析并总结了改进方法。

Abstract: The ResNet and EfficientNet are used to classify the remote sensing image. Among ResNet and Effi-cientNet, the former uses an 18-layer network structure and has a classification accuracy of 59.8% in the test set; the latter uses EfficientNet-B0 and EfficientNet-B1, with the highest classification accuracy of 92.6%. After that, we replace the SGD optimizer with Ranger optimizer. Although the new classification accuracy is similar to the original one, using the Ranger optimizer to train the EfficientNet model has stronger stability and faster training speed. Finally, we analyzed the test results, listed the classification accuracy of each type of scene, analyzed and summarized the improvement methods.