High-resolution remote sensing image classification method based on residual network and transfer learning

Abstract

The invention belongs to the technical field of image processing and analysis, and particularly relates to a high-resolution remote sensing image classification method based on a residual network and transfer learning. The method comprises the following steps: establishing a target data set, and carrying out label labeling based on a ground feature category of the target data set; constructing a reconstruction data set and a test set only containing the high-resolution remote sensing images of the determined categories; improving the Resnet101 deep residual error network; determining a deep residual network training model and parameters; obtaining a multi-scale scene classification and voting result of the target data set; obtaining target data set scene classification results of multiple scales; completing the pre-training of the model, and obtaining a multi-scale scene classification result of the target data set; and voting and regenerating all high-resolution remote sensing images to finish a classification task. According to the method provided by the invention, the improved Resnet101 deep residual network is utilized to perform feature extraction and pre-training of the source domain data set, multi-scale scene classification is performed, and the overall precision can reach more than 95%.

Description

technical field [0001] The invention belongs to the technical field of image processing and analysis, and in particular relates to a high-scoring remote sensing image classification method based on residual network and migration learning. Background technique [0002] High-resolution remote sensing images (or high-resolution remote sensing images) can observe the earth from meter-level or even sub-meter-level spatial resolution, and clearly express the spatial structure and surface texture characteristics of ground objects, so as to distinguish the inner surface of objects. The detailed composition provides ideas for effective geoscience analysis. Compared with natural images, high-resolution remote sensing images not only have rich spatial, shape and texture features. Therefore, the use of high-resolution remote sensing images for scene classification is of great significance for image interpretation and real-world understanding. With the rapid development of high-resolut...

AI Technical Summary

Problems solved by technology

[0003] 1) High-resolution remote sensing images have rich spatial information. Existing methods usually only extract deep high-level semantic features while ignoring the underlying detailed features, making it difficult to fully utilize the fine spatial information of high-resolution remote sensing images. The classification effect of small-scale ground objects and scenes is not good

[0004] 2) It is difficult to label high-resolution remote sensing images. The existing methods generally use the public dataset UC Merced LandUse Dataset (the image pixel size is 256x256, and it contains 21 types of scene images, each type has 100 images, a total of 2100 images), WHU- RS19 Data Set (image pixel size is 600x600, contains 19 types of scene images, each type has about 50 images, a total of 1005) and AID dataset (image pixel size is 600x600, contains 30 types of scene images, each type has 200 -400, a total of 10,000), etc. as the original training data set for migration learning

However, the above-mentioned datasets are too finely labeled, the samples of each category of image labeling are small, and the scene classification results are complicated, which is difficult to meet the needs of actual production and life for scene classification.

Images