Research On Cropland Recognition And Extraction Technology Of High-Resolution Remote Sensing Images Based On Convolutional Neural Networks

The cropland is the basis of agricultural production activities,and the accurate and rapid extraction of cropland has important significance for monitoring agricultural resources and ensuring food security.Using remote sensing image classification is a fast and effective method for cropland extraction.The ground objects in high-resolution remote sensing image data are rich in details,and can be used to extract smaller cropland parcels.However,due to the variety of crops planted in cropland,the spectrum and texture characteristics of different plots of cropland in high-resolution remote sensing images are quite different,and the phenomenon of same objects with different spectrums is serious,which makes the extraction of cropland facing great challenges.Traditional remote sensing image classification methods based on machine learning cannot extract deep features of cropland from images and require manual selection of features,so they cannot accurately extract cropland parcels.The convolutional neural network in deep learning is one of the most popular methods in image processing.It can automatically learn and extract deep features from images,greatly improving the accuracy of image processing tasks.Therefore,this article applies the convolutional neural network method to the task of identifying and extracting cropland from highresolution remote sensing images,and proposes a method for extracting cropland based on convolutional neural network.The main research contents and conclusions are as follows:（1）The cropland remote sensing recognition based on convolutional neural network of transfer learning.The VGGNet,the Res Net,the Google Net,the Xception,the Dense Net and the HRNet convolutional neural network models were used for pretraining on the NWPU-RESISC45 remote sensing scene classification data set.Then transferred the pre-trained weights to the GF-1 remote sensing images for training,and obtained the cropland recognition results of these models.Finally,the experiment compared the classification and the cropland recognition results of different models using transfer learning and not using transfer learning under 50% training data and 80%training data.The research results showed that in the case of a small amount of data,the use of transfer learning could effectively improve the accuracy of the remote sensing classification and the cropland remote sensing recognition of convolutional neural network models.And the F1-score of cropland recognition was between 0.89 and 0.96.（2）The CMHRNet model construction and cropland extraction from Gaofen-1remote sensing images.In view of the fact that most neural network models would cause information loss in the extraction of cropland from remote sensing images,the lightweight high-resolution network HRNet was used to obtain more accurate spatial information,and the class attention mechanism module and the multi-parallel atrous spatial pyramid pooling module were used to better combine global and local context information to improve the HRNet model,and build an improved HRNet model named CMHRNet.And through experiments to compare the results of cropland extraction and classification on the GF-1 remote sensing images of the CMHRNet,the HRNet and the UNet models,to verify the effectiveness of the CMHRNet model.Finally,the transfer learning experiment of cropland recognition was used to verify whether transfer learning could further improve the cropland extraction accuracy of the CMHRNet model.The experimental results showed that the CMHRNet had learned more detailed information in remote sensing images than the HRNet and the UNet,and could extract cropland more completely,with richer details and clearer edges,which improved the accuracy and precision of ccropland extraction and classification.Among them,the overall accuracy of cropland extraction reached 0.9474,the F1-score reached 0.9118,the Kappa coefficient reached 0.8237,the MIOU reached 0.8428,the accuracy of cropland reached 0.88,the overall classification accuracy also reached 0.8914,and the amounts of parameters were 63.4% less than the UNet.And transfer learning based on the remote sensing scene classification data set could further improve the cropland extraction accuracy of the CMHRNet model.