| Large-scale,rapid and accurate identification of rice planting areas is of great significance to national agricultural resource monitoring and national food security.With the continuous improvement of remote sensing real-time and operability,extracting natural resource information through remote sensing image interpretation has become a very efficient way.Traditional remote sensing interpretation methods based on pixels and objects have shortcomings such as shallow features,high degree of manual intervention of parameters,and inaccurate extraction of multiple features.With the continuous emergence of deep learning theory and technology represented by convolutional neural networks,it has achieved good results in the field of image processing.This paper uses high-scoring remote sensing images as the data source and uses the convolutional neural network method to carry out relevant research work on rice identification.The main contents are as follows:1.Through the research and exploration of deep learning methods and the characteristics of common remote sensing feature classification models,a deep learning model based on the spatial and spectral features of GF-2 remote sensing images is constructed;the purpose of machine autonomous learning and automatic extraction of rice pixels at the pixel level is realized.The model uses an improved U-Net semantic segmentation model,including two core modules,Encoder and Decoder.Among them,the Encoder module refers to the Res Net-18 model,expands the receptive field through the maximum pooling operation,and uses the residual convolution block to extract image features of different scales and dimensions;the Decoder module is realized by four upsampling of the U-Net model;2.Expand the rice sample size through the data enhancement algorithm,build a rice booting stage sample database using GF-2 image as the data source,adjust the model parameters to obtain an optimized model,and finally use the rice tags for model training,verification and testing;3.In model recognition,data input and result output cropping strategies are proposed to realize model recognition of input images at any scale;4.Finally,the impact of different sample sizes and different classifications of ground features on the accuracy of the model is explored,and compared with traditional object-oriented and support vector machine algorithms.Through experimental verification,compared with the traditional algorithm,the model in this paper makes full use of the spatial and spectral feature details of the image,realizes the autonomous learning of rice features,and improves the accuracy and generalization ability of rice identification.The method in this paper effectively realizes the in-depth exploration of the value of surveying,mapping and mapping results,and provides technical support and reference for the comprehensive utilization of high-resolution images and natural resource results to identify rice and other ground features. |
