关键词: 冬小麦, 无人机, 植被指数, 生育时期, 特征筛选, 氮效率分类

Abstract:

【Objective】To explore the potential of UAV remote sensing in nitrogen efficiency classification and recognition, a nitrogen efficiency classification method for wheat varieties was constructed, so as to provide the theoretical basis and technical support for nitrogen efficient variety screening.【Method】Six agronomic indicators related to nitrogen efficiency at maturity stage (yield, plant nitrogen accumulation, nitrogen physiological use efficiency, plant dry biomass, total nitrogen uptake of grains, and N harvest index) were used to construct the principal component synthesis value, and K-Means cluster analysis was performed on them. The 121 wheat varieties were divided into three types: high, medium, and low nitrogen efficiency types. A UAV remote sensing platform equipped with a multi-spectral camera was used to obtain remote sensing images of wheat at the jointing, booting and flowering stages, and 34 vegetation indices were extracted to analyze the correlation between vegetation index and nitrogen efficiency comprehensive value. The accuracy of nitrogen efficiency classification models of support vector machine (SVM), random forest (RF), and K-nearest neighbor (KNN) classification methods were compared, and the overall classification accuracy (OA) and Kappa coefficient were used to compare the classification and recognition ability of wheat varieties in different growth periods. Three different feature set screening methods(ReliefF algorithm, Boruta algorithm and RF-RFE algorithm) were used to comprehensively evaluate the optimized feature subsets, and an appropriate classification and recognition method for wheat varieties nitrogen efficiency was established.【Result】With the progress of wheat growth stage, the correlation between vegetation index and the comprehensive value of nitrogen efficiency gradually increased, which reached the highest correlation coefficient at flowering stage (r=0.502). The full feature set of vegetation indices was used to classify the nitrogen efficiency of wheat varieties. For the data of single growth stage, SVM model had the best classification accuracy at flowering stage (OA=77.1%, Kappa=0.591), and the worst classification accuracy at jointing stage (OA=65.6%, Kappa=0.406). In general, the classification accuracy of nitrogen efficiency of varieties with multi-growth stage data fusion was higher than that of single growth stage, among which SVM model with jointing stage + booting stage + flowering stage had the best classification accuracy (OA=80.6%, Kappa=0.669). In order to reduce the number of feature set variables in multi-growth period data fusion, the feature optimization effects of RF-RFE, Boruta and ReliefF algorithms were compared and analyzed. The optimal feature subset based on RF-RFE algorithm had the highest classification accuracy, and its OA and Kappa coefficients were 4.0% and 10.1% higher than those of the full feature set classification model, respectively. Among them, the data fusion of three growth stages had the best classification accuracy (OA=85.4%, Kappa=0.749).【Conclusion】The nitrogen efficiency evaluation method with six nitrogen efficiency indexes - principal component analysis -K-Means were established in this study. The RF-RFE algorithm effectively optimized the number of characteristic subsets of the multi-growth period combination, and obtained high classification accuracy. A nitrogen efficiency classification model of wheat varieties based on the fusion of multi-growth period combination and RF-RFE-SVM technology was established, which provided the theoretical basis and technical support for the rapid and accurate classification and identification of wheat varieties with nitrogen efficiency.

Key words: winter wheat, UAV, vegetation index, multiple growth periods, feature selection, nitrogen efficiency classification