The Full Convolutional Neural Network Deep Learning Based Forest Classification Method For High Spatial Resolution Satellite Remote Sensing Image

In recent years,deep learning has made great progress in face recognition,image segmentation and speech recognition.The full convolutional neural network can perform automatic feature learning on multidimensional data,and is successfully applied to the semantic segmentation of images.At the same time,the fast development of high resolution remote sensing satellites provides advantages and possibilities for imporving forest resources monitoring techniques.In various network models of deep learning,the full convolutional neural network U-net can effectively improve the effects of remote sensing target recognition and feature classification.At the same time,existing studies have also shown that the classification effect of FCN-8s deep learning model can be optimized by adjusting a small number of relevant remote sensing features and using CRF post-processing method.However,whether this strategy can also be applied to U-net forest classification has been less reported.Aiming at meeting the application requirement of national forest resource planning and design survey for high resolution multi-spectral remote sensing classification technique,this study takes Wangyadian Forest Farm in Chifeng City of Inner Mongolia Autonomous Region as the research area,and carries out the research on deep learning forest classification method based on full-convolution neural network for space-borne multi-spectral-panchromatic remote sensing data with high spatial resolution characteristics.The main research contents and conclusions are as follows:(1)Full convolution network for forest type classification using multi-spectral remote sensing imageUsing the GF-2 multi-spectral image in winter of 2015(without panchromatic image)and inputing the same remote sensing data,training samples and test samples,two deep learning classification methods(U-net,FCN-8s)and two traditional machine learning classification methods(SVM,RF)were compared and evaluated in order to develop a classification method with relative best classification performance.A deep learning classification method based onfull convolution network for forest type classification using multi-spectral images with relatively optimal classification performance is developed,and this optimized U-net model is named as U-net-NDVI-CRF.The experimental results show that: The overall classification accuracy of the optimized U-net model is 84.89%,and the Kappa coefficient is 0.82,which is higher than that of the U-net model without NDVI feature and U-net model without CRF post-processing;Compared with the classification results of FCN-8s,SVM and RF using the same strategy,the classification accuracy of the optimized U-net model is greatly improved.(2)Full convolution network for more detailed forest type classification using Pan-sharpen multi-spectral remote sensing image Considering that winter images are more suitable for the classification of deciduous and non-deciduous forest types,but not suitable for classifying different types of deciduous forests,while images of growing season have the advantage of identifying more forest types,in this study,multi-spectral and panchromatic images of GF-2 from the 2017 growth season were used to do in-depth study on the applicability of the U-net-NDVI-CRF deep learning classification method.On the one hand,the applicability of this method to images acquired in different seasons is explored;on the other hand,the applicability of this method to Pan-sharpen multi-spectral images and increased number of categories to be classified is also explored(12 categories in total,including 6categories of forest types).The results show that the number of distinguishable forest types can be effectively increased(from 3 to 6)by combining multi-spectral and panchromatic bands and then using U-net-NDVI-CRF for the classification of more forest types.Moreover,this method has higher classification accuracy than the other two kinds of FCN methods to be compared with.In summary,this study proposes a classification method based on U-net model of full convolutional neural networks: U-net-NDVI-CRF.It not only has good adaptability to GF-2remote sensing images acquired in different seasons,but also achieved relatively best classification results for both identifying few categories using only multi-spectral images and more classes based on Pan-sharpen multi-spectral images.The proposed method is suitable fordetailed forest type classification using high spatial resolution multi-spectral satellite remote sensing images.