Crop Classification With GF-6 WFV Imagery Based On Object-oriented Analyses And Deep Learning

As the basic supporting data,satellite images of remote sensing have huge demand and become increasingly important.With the construction of China's High-resolution Earth Observation System for Major Scientific and Technological Projects,the research and application of acquired remote sensing images with multiple spatial,temporal and spectral resolutions in land cover utilization,resource and environment monitoring and other fields need to be continuously promoted.Among them,satellite of GF-6 is China's first remote sensing satellite with red-edge band,and it is also the world's first wide-view multi-spectral high-resolution satellite.The research and utilization of wide-area satellite image data have not been fully developed,and the research and utilization of its rich spectral information will be of great significance.This article analyzes and summarizes the advantages and disadvantages of current remote sensing image classification applications,and adopts the idea of hierarchical classification to conduct crops classification for remote sensing images according to the characteristics of GF-6 WFV images.By combining object-oriented and constructing a new convolutional neural network model(RE-CNN)proposes a crops classification method for remote sensing images suitable for the GF-6 WFV red-edge bands.First,multi-scale image segmentation and ESP tools are used to perform multiple image segmentation experiments,the best segmentation parameter is selected by combining Trial and Error method and manual visual discrimination to further complete the image segmentation.Select the training samples and construct vegetation index for the newly added red-bands in the image,and complete the CART decision tree training based on the characteristics of the image object.The object-oriented CART decision tree was used to eliminate the salt and pepper phenomenon and the vegetation area images were extracted.Based on an in-depth analysis of the structural principles of deep convolutional neural networks and the difficulties and key technologies in model training,this paper builds and trains a convolutional neural network model(RE-CNN)based on Python and Numpy libraries,and converts the extracted vegetation area image data into input data of deep convolutional neural network.Based on the newly added red-edge bands in the image,the classification control experiment with or without red-edge bands is set up,and through the constructed CNN model to complete the GF-6 WFV image crop classification.The experimental results show that the object-oriented classification method can effectively avoid the salt and pepper phenomenon,and at the same time obtain a good classification effect for the vegetation extraction in the study area.In the control experiment with or without the red-edge bands,the classification of the crop classification of the red-edge bands group has achieved good results,and the overall accuracy is as high as 94.38%.Compared with another group without red-edge bands,the classification accuracy is improved by 2.83%,which effectively validates the effectiveness of the RE-CNN model for GF-6 WFV image crop classification and the sensitivity of the newly added red-edge bands to crop classification recognition.Important theoretical support and technical references are provided for the research expansion of GF-6 WFV images for fine crop classification.