CNN-based Land Cover Classification Research

In recent years,with the rapid development of satellite sensors and aerial photography technology,the acquisition methods of high spatial resolution remote sensing images have become increasingly mature.The high resolution image data provides geometric details and rich texture features of the land cover type.While heterogeneous similarities and homogeneous heterogeneity on high resolution images also increase the complexity of feature extraction.The traditional methods of object-based image analysis(OBIA)are difficult to solve the problems mentioned above in high-resolution images.Convolutional Neural Network(CNN),along with the advent of the era of deep learning,provides the possibility to extract depth features hidden in high resolution images.The CNN contains multiple hidden layers,which make it highly robust in image classification.At present,some scholars have proposed pixel-based CNN(Pixel-CNN)method for very high resolution image classification and achieved higher classification accuracy than that of general classifiers.However,there are still problems such as massive spatial processing units,huge computational time consuming,pixel-level salt and pepper phenomenon.In order to solve the these problems,this paper proposed a high resolution image remote sensing classification method combining OBIA and CNN.Based on the requirements of the influence block input,a super-pixel based CNN(SEEDS-CNN)is proposed.SEEDS-CNN combines Superpixels Extracted via Energy-Driven Sampling(SEEDS)with CNN: the SEEDS method is used for image segmentation,and the superpixel center point is used as the classification point in classification.SEEDS-CNN achieves higher classification results than the traditional OBIA method.At the same time,compared with the traditional Pixel-CNN,the classification efficiency of the image is greatly improved without losing the classification accuracy.However,SEEDS-CNN still has the problems of poor classification of ground objects,super-pixel level salt and pepper errors and more space processing units.Therefore,for the above problems of SEEDS-CNN,this paper further proposed a region-based majority voting CNN(RMV-CNN)method.RMV-CNN combines the segmentation method with better object boundary segmentation effect with CNN: multi-scale segmentation algorithm is used for segmentation,multiple samples to be classified are generated in each segmented region,and then these are to be classified.The samples are classified by pre-trained CNN model,and finally the majority voting in their respective regions is used to obtain the final classification results.The experimental results show that:(1)OBIA-CNN classification efficiency is much higher than that of Pixel-CNN;(2)RMV-CNN classification accuracy is about 2% higher than that of SEEDS-CNN and Pixel-CNN;(3)RMV-CNN The boundary classification works well.At the same time,the paper analyzes and discusses the scale effect of high resolution image classification.After multi-scale series experiments,the scale effect of classification is as follows:(1)Multi-scale classification results are significantly better than single-scale classification results;(2)RMV-CNN solves the salt and pepper phenomenon of high-resolution image classification at a small scale to some extent;(3)In the extreme case,the small-scale classification results are the best.