Crop type recognition based on GF-2 images and Sentinel-2 images

【Objective】Based on fused images of GF-2 PAN and Sentinel-2 MSI， this study proposed a stepwise hie-rarchical extraction strategy for crop type recognition， aiming to provide technical support for crop classification in complex planting areas.【Method】The main planting areas of Yanjiang Street， Qilin District， Qujing City， Yunnan Province were taken as the study area to acquire data of GF-2 MSI， GF-2 PAN， and Sentinel-2 MSI， and the original data was fused using different fusion methods. The crop planting areas were extracted based on a decision tree. The optimal classification configuration was identified through classification band selection， vegetation index selection， texture feature optimization based on ReliefF， and optimal segmentation scale determination based on ESP2； the object-based classification of extracted planting areas was performed using， support vector machine （SVM）， random forest （RF）， and nearest neighbor algorithm.【Result】For comprehensive evaluation of the fused images， the fused images acquired using the PCA fusion method， db3 wavelet method， Sym3 wavelet method， and Coif5 wavelet method were used to classify crops. Under the optimal classification feature configuration （DVI + Red-3 + B-information entropy + NIR-homogeneity + G-homo-geneity + B-homogeneity + SWIR2-homogeneity + R-homogeneity + SWIR1-homogeneity + G-information entropy + SWIR2-information entropy）， and using eCognition 9.0 software， SVM， RF， and nearest neighbor algorithm were employed to classify different crops in the study areas. The Coif5 wavelet method with the RF algorithm achieved the highest overall classification accuracy （90.42%）， with a Kappa coefficient of 0.8739. Among different crops， high classification accuracy was found in grapes and other ground objects， followed by that for vegetables， while the accuracy for wheat and legumes was relatively low.【Conclusion】The images fused using the Coif5 wavelet method are the best. The fused images are combined with BSI， EIBI， WI2015， and wetness component of KT transform to construct a decision tree， and integrating ReliefF + DVI + Red-3 with a RF classifier yields an object-based classification model with strong performance for different crops. Texture， spectral， and band features calculated based on fused images can recognize crops in complex planting areas； masking large expanses of irrelevant ground objects can reduce potential misclassification， and introducing red edge bands and DVI further enhances the advantages of fused images in recognizing different crops.