Objectives Pyramid scene parsing net (PSP Net), a kind of neural network with deep structure, extracts the features of remote sensing image better than the traditional model, e.g. artificial neural network, and supports vector machine. This algorithm can embed the contextual features of difficult scenes into the pixel prediction framework of the full convolutional neural network (FCN), fully understand the scene, realize accurate prediction of each pixel category, location and shape, and fuse local and global information together to propose an optimization strategy for moderate supervised loss. A deep learning algorithm based on PSP Net is proposed to achieve higher accuracy in image classification and effectively promote remote sensing image automation and intelligent interpretation.
Methods Conventional architecture for fast feature embedding (CAFFE) is a deep learning framework with is expressive, fast and thought modular, and supports a variety of deep learning architectures for image classification and segmentation. PSP Net model under the CAFFE framework is used by modifying and optimizing the network to eliminate the overfitting effect. Using the remote sensing image of Hubei Province as the experimental data, the land with the 30 m resolution is studied with the aid of the analysis ability of the context scene of PSP Net. The Python program is used in the CAFFE depth learning framework for operation. The operation network obtains the global feature information through the core structure (pyramid pooling module) and completes the generation of the data set. In the experiment, the 507 standard partial images of 900×600 pixels in the landsat image of Hubei Province are used, and the sample sets suitable for depth learning are generated by pre-processing. Among these, 300 data are selected, including 240 training sets, 44 prediction sets, and 16 validation sets. The overall accuracy (OA) index is used as the prediction accuracy index of the preliminary evaluation model to evaluate the accuracy of the prediction model.
Results PSP Net model under the deep learning framework of CAFFE is used to train the sample data. The learning rate of 10×10-10 is set, and the million times training model is selected to protect against overfitting of the data. Through the generalization of the model and the generalization and iteration of the samples, the land cover of the 3 phases of TM images about Hubei Province are classified, and the classification accuracy is 82.2%, 83.4% and 83.7%, respectively.
Conclusions The results show that the land cover classification of remote sensing images can be realized quickly, effectively and accurately by PSP Net.