With the continuous progress and development of human society in nowadays, land-use pattern gets its diversification. However, due to the variety of land-use patterns, it is impractical for people to precisely label the land-use pattern of somewhere. Remote sensing images, recording the characteristic of terrestrial objects in different height and angle, could be divided into three classes: marine remote sensing, terrene remote sensing and meteorological remote sensing. Recording in different height and angle results in difference in images resolution, which makes artificial classification in remote sensing images more difficult. Meanwhile, the development of aeronautical technology and satellite provide us with huge original data set which could be utilized to observe and measure the terrestrial structure. Besides, understanding remote sensing images is vital in land using, city planning, protection of ecological resources and so on.
Deep learning, a subfield of machine learning that uses deep neural networks, has achieved state-of-art results in field such as images because of its successive layer process, internal changes of characteristics and sufficient model complexity. Convolutional neural network is one of the most appropriate neural networks in computer vision application, it is capable of using local operation to separate and abstract characterization. Because the pixel in the adjacent area of remote sensing images are highly correlated, convolutional neural network can utilize the 2-Dimension structure of images without using the connection peer to peer between all single pixel.
The problem discussed in this project is classification of remote sensing images based on convolutional neural networks, which belongs to multi-classification problem of supervised learning owing to the diversity of remote sensing images. Firstly, build the basic unit of neural network, then expand the depth of network in order to learn much characteristic of remote sensing images with combination of batch normalization and dropout to improve the accuracy of classification.