Research On Crop Classification Algorithms Based On Deep Learning Using High-resolution Remote Sensing Imagery

In recent years,with the increase in satellites at different spatial,temporal,and spectral resolutions,remote sensing techniques have emerged as optimal tools to identify crop types over larger areas.Timely and accurate crop-type classification is essential for estimating crop yields,strengthening crop production management,and crop insurance.How to efficiently exploit useful information from remote-sensing imagery for better earth observation is an interesting but challenging problem.This paper summarizes the research results of crop classification algorithms based on deep learning and high-resolution remote sensing images at home and abroad.In order to meet the demand for crop classification information in Jilin Province,Northeast China,this paper conducts in-depth exploration of crop classification methods.The specific research contents and innovation results are as follows:（1）Research on crop classification of optical remote sensing images based on deep learning.This study used multi-temporal Sentinel-2 and GF-2 as data sources,and selected Jilin Province in Northeast China as the study area to classify the typical crop types（rice,corn,soybeans）in the area.This experiment used CNN and VGG as classification methods to explore the influence of the addition of red edge bands,vegetation index,etc.on crop recognition results.The experimental results showed that the addition of vegetation index and red edge band improved the recognition of crops,and the VGG network performed best in crop recognition accuracy,with an overall classification accuracy of 94.89% and a Kappa coefficient of 0.925.（2）Research on crop classification based on optimal feature selection and hybrid CNN-RF networks.To address the challenge of the large number of features and the information redundancy between features,this paper proposed an optimal feature selection method（OFSM）and compared with two traditional feature selection methods（TFSM）: random forest feature importance selection（RF-FI）and random forest recursive feature elimination（RF-RFE）.Although the time required for OFSM was26.05 s,which was between RF-FI with 1.97 s and RF-RFE with 132.54 s,OFSM outperformed RF-FI and RF-RFE in terms of the overall accuracy（OA）of crop classification by 4% and 0.3%,respectively.Moreover,this paper designed two hybrid CNN-RF networks（Conv1D-RF and VGG-RF）and used the optimal features selected by OFSM to compare the four networks: Conv1D-RF,VGG-RF,Conv1 D,and VGG.Conv1D-RF achieved the highest OA at 94.27% as compared with VGG-RF（93.23%）,Conv1D（92.59%）,and VGG（91.89%）,indicating that the Conv1D-RF method with optimal feature input provides an effective and efficient multi-temporal crop-type classification method.（3）Research on fully automated classification method for crops based on spatiotemporal deep-learning fusion technology.First,realize the automatic selection of training samples,including pixel samples and image-block samples.Secondly,this paper proposed Geo-3D CNN and Geo-Conv1 D with geographic information.Finally,we used the active learning strategy to integrate the classification advantages of Geo-3D CNN and Geo-Conv1 D to further improve the identification accuracy of crops in the study area.Experimental results showed that the proposed fully automated sample selection method can ensure that a large number of reliable samples were selected and provide a guarantee for accurate crop identification.The crop classification method based on spatio-temporal deep learning fusion technology can achieve the highest OA at 93.18%,slightly higher than that of Geo-Conv1D（92.07%）and much higher than that of Geo-3D CNN（90.67%）,indicating that the proposed classification fusion method is effective and efficient in multi-temporal crop classification.This paper adopted multi-temporal optical remote sensing imagery as the data source.First,deep learning technology and multi-temporal optical remote sensing imagery were used to realize the recognition of crops.In order to improve the crop mapping,this paper proposed the optimal feature selection method and designed hybrid CNN-RF networks.Finally,a fully automated training sample selection method and a novel crop classification method based on spatio-temporal deep learning fusion technology were proposed.This study provides a reference for future crop classification research in Northeast China.