A New Convolutional Neural Networks for Land Use Classification.

In the land cover classification research, traditional remote sensing image land cover classification methods are purely based on the spectral information of ground objects, which means they do not make full use of the spatial information of ground objects. So it cannot achieve satisfactory classification results. Moreover, the classical convolutional neural network model is not suitable for multispectral image data. To solve these problems, this paper proposes a new method for land cover classification based on our own new convolutional neural network. First, we apply MNF to multispectral image data to discard noise and achieve dimensionality reduction; Next, we decompose the multispectral image data after dimensionality reduction into patch for each pixel; Finally, the proposed CNN is used to extract LULC classification information. In this paper, Landsat-8 data is used to analyze land use/cover and ecological environment in Ganji-ngzi District of Dalian City with the above method. The overall accuracy of the method in this paper can reach 93.4%, 9 percentage points higher than SVM. Experimental results show that the method with CNN framework proposed in this paper effectively combine spectral and spatial information and achieve excellent classification results. It is a feasible method to obtain land use/cover classification information accurately, which can provide reference for land cover classification research.