Study On Typical Crops Classification With High-resolution Remote Sensing Images Based On CNN

As a strategic natural resource,cultivated land is the material basis and important prerequisite to ensure the safety of food production in China.How to quickly and accurately obtain the spatial distribution characteristics of the cropland,find out the types of crops planted,at the same time master the state of cultivation is one of the key task of government departments.With the development of remote sensing technology,high-resolution visible light remote sensing imagery has become a convenient and reliable source of remote sensing data.At present,the use of such remote sensing images to carry out large-scale typical crop classification research is of great practical significance to promote agricultural modernization and maintain regional ecological stability and sustainable development.Using traditional classification methods for remote sensing information extraction often fails to make good use of the high-dimensional features in the images,and it is difficult to obtain ideal classification results.Convolutional neural network,as an emerging technology in the field of image recognition in recent years,can automatically learn the deep features of images to make accurate classification decisions,which brings new opportunities for better high-resolution remote sensing image classification results.In this context,The research work carried out in this paper and related results are as follows:（1）Aiming at the problems of low spectral information in crop high-resolution remote sensing images,insufficient utilization of texture features,low recognition accuracy,and severe salt and pepper noise,deep learning techniques are used for crop classification and recognition.The composite-wing UAV,the GF₁ satellite high-altitude remote sensing platform,and remote sensing imaging equipment were selected to obtain remote sensing data in the Shawan and Yuili counties.Based on the acquired high-resolution remote sensing data characteristics and local crop phenology information,a convolutional neural network was designed to extract the fine type information of the crops.（2）Based on the UAV data,the parameters of the training process are optimized into different groups by adjusting the network parameter settings and the sample spectrum combination,which are the learning rate-read batch adjustment group,sample spectral feature adjustment group,and kernel scale adjustment group discussed the influence of the adjustment of related parameters on the accuracy of crop classification of convolutional neural networks.The research show that the convolutional neural network can effectively extract the crop information in the high-resolution remote sensing image of the drone.Except for the sparse and mixed crops at the edge of the plot,there will be a little mis-segmentation.The optimized network model can achieve an overall classification accuracy of 98% for three typical crops.During training,the adjustment of model parameters will affect the final training result of the model.For typical crop samples in highresolution visible light remote sensing images with high density,small features,and rich texture,a large learning rate（0.1）,small The convolution kernel（7 × 7）and the appropriate deep layer（7 layers）of the network model are used for feature extraction and classification,which accelerates the convergence of the network accuracy while ensuring that the extraction of small features in the sample is not missed.The samples containing different combinations of spectral features will also affect the training of the network model.In the visible light band,the blue light band contains the green light and red light band samples,which is more conducive to training the convolution god to recognize the network for crops.,When the sample contains three bands at the same time,it will bring higher training recognition accuracy and more stable recognition effect.（3）Taking the satellite remote sensing data as the object,U-Net model was selected to extract the farmland space and its category information.In order to make full use of the phenological characteristics of the study area and improve the recognition effect,the vegetation index NDVI value and texture feature gray were combined with the winter and summer remote sensing images.The energy value of the symbiosis matrix is used as a characteristic band,and the network is added for training to obtain U-Net recognition models of farmland（cotton-covered arable land）,fruit forest,and uncultivated farmland,respectively.The comprehensive model recognition results and judgment conditions finally determine the coverage type of the study area.For high-resolution satellite remote sensing data,the U-Net model can better extract farmland information,and the recognition accuracy of cotton-covered farmland in the study area can reach 90.83%.After adding winter data,vegetation index,and texture feature bands for training,the model can effectively improve the recognition effect of fruit forest cultivated land and uncultivated cultivated land.The overall recognition accuracy is 88.39% and 79.51%,which are 4.67% and 6.11% higher than traditional methods.With the help of discrimination conditions and winter and summer data,it helps to increase the precision of extracting intercropped arable land.At the same time,compared with traditional classification algorithms such as support vector machines and random forest classifiers,this method can reduce the "pepper and salt noise" among the same crops in the plot "To avoid large-scale misallocation of uncultivated land.