面向无人机变量施肥的寒地水稻叶绿素高光谱反演方法研究OA

Unmanned aerial vehicle(UAV)-based fertilization is increasingly applied in rice production;however,fertilization prescriptions are still largely dependent on empirical knowledge and lack decision support based on real-time crop nutrient status.Meanwhile,limitations in hyperspectral data processing and model construction restrict the high-precision estimation of chlorophyll content in cold-region rice,thereby constraining the accurate implementation of variable-rate fertilization.To address these issues,this study combined multiple spectral preprocessing methods with feature variable selection techniques,including uninformative variable elimination(UVE),competitive adaptive reweighted sampling(CARS),iterative retention of informative variables(IRIV),and variable iterative space shrinkage approach(VISSA),to extract key spectral bands.Models including least squares support vector regression(LSSVR),particle swarm optimization-based LSSVR(PSO-LSSVR),partial least squares regression(PLSR),and a deep learning model(1DCNN-WN)were developed for chlorophyll content inversion.Based on the inversion results,nitrogen fertilization prescription maps were generated to guide UAV-based variable-rate fertilization at the tillering stage.The results showed that the SNV-IRIV-1DCNN-WN model achieved the highest inversion accuracy under the optimal preprocessing and feature selection combinations,with an R2 of 0.687,which was 0.047,0.031,and 0.027 higher than those of SNV-UVE-LSSVR,NL-VISSA-PSO-LSSVR,and MSC-IRIV-PLSR,respectively,while the RMSE was reduced by 0.281,0.147,and 0.107.The coefficients of variation for no fertilization(N0),low fertilization(N1),standard fertilization(N2),and excessive fertilization(N3)treatments were 6.08％,4.16％,3.42％,and 3.07％,respectively,indicating good fertilization uniformity.These findings provide theoretical support and technical guidance for hyperspectral-based chlorophyll inversion and precision fertilization in cold-region rice production.