Crop recognition and evaluationusing red edge features of GF-6 satellite

Abstract

The application of red edge features, which are sensitive bands of vegetation, is a high-technology method for remote sensing to identify crops and realize precision agriculture. Multispectral GF-6 image of the study area in the northern region of Songnen Plain in Heilongjiang Province pioneers the use red edge bands in China. A total of 82859 crop samples of corn, soybean, and rice were used as research objects. The classification accuracy of crops was evaluated and the performance of red edge features in crop identification, such as red edge bands and vegetation index, was discussed from the following aspects. (1) Statistical characteristics of radiance values of crop samples initially showed that discrimination is better at Band 5-0.710 μm and Band 6-0.750 μm in the two red edge bands than findings of other GF-6 bands. (2) Traditional normalized (NDVI) and red-edge normalized difference vegetation indexes, namely, NDVI710 and NDVI750, are constructed. Results showed that the two indexes are more significant than the traditional NDVI in the classification of crop samples characterized using J-M distance. (3) Effective bands are screened using various methods, and classification strategies for the four types of crops are formulated using Support Vector Machine (SVM). Crop classification in the study area is completed using five sets of random sample segmentation schemes, namely, 5: 5, 6: 4, 7: 3, 8: 2, and 9: 1. Twenty types of classification accuracy demonstrated a kappa coefficient higher than 0.9609 and overall accuracy higher than 0.9742. The 5: 5 and 8: 2 segmentation schemes in the column direction exhibited the highest and lowest accuracy, respectively. The sorting accuracy in the horizontal direction demonstrated the following order: SVM-RFE > SVM-RF > SVM with red edge bands > SVM without red edge band, which also showed that the participation of red edge vegetation index and red edge band significantly improves the recognition precision of crops. (4) SVM-RFE and SVM-RF both obtained minimal misclassifications due to the lack of other samples, such as waters. However, SVM-RFE is superior to SVM-RF in terms of classification accuracy and image detail display with a kappa coefficient and overall accuracy of 0.8060 and 0.8743, respectively, in the cross-validation of two classified images. Hence, the red edge feature of GF-6 is superior in crop recognition with its significantly improved recognition accuracy. Subsequent investigations can focus on developing additional red edge-related vegetation indexes and optimize the role of red edge characteristics in precision agriculture.