基于高光谱技术的设施黄瓜叶片氮含量快速检测研究OA

[Objective]Nitrogen is a core macronutrient that drives the growth,development,yield formation,and quality establishment of greenhouse cucumbers.Real-time monitoring of leaf nitrogen concentration is crucial for precise fertilization and green production.This study aims to rapidly and non-destructively monitor leaf nitrogen concentration in greenhouse cucumbers using hyperspectral imaging systems,with the goal of improving agricultural production efficiency and accuracy,and achieving intelligent management and precise fertilization.[Method]The experiment was conducted in a second-generation solar greenhouse in Helan County,Ningxia.Six nitrogen application levels(N0-N5)were set,and the cucumber cultivar"De'er 15"was used.First,hyperspectral images of cucumber leaves during the flowering and fruiting periods were collected,and leaf nitrogen concentration was precisely determined through laboratory chemical analysis to obtain baseline modeling data.Next,samples were divided into training and prediction sets in a 4∶1 ratio using the SPXY algorithm.The raw spectral data were preprocessed using the moving average(MA)and Savitzky-Golay(SG)methods.Feature wavelengths were then extracted using competitive adaptive reweighted sampling(CARS),uninformative variable elimination(UVE),and a combination of UVE and CARS(UVE-CARS).Finally,three machine learning methods—random forest(RF),extreme learning machine(ELM),and convolutional neural network(CNN)—were used to establish prediction models for leaf nitrogen concentration during the flowering and fruiting stages of cucumbers.[Result]Preprocessing of the raw spectral data effectively improved the accuracy of the prediction models,with the SG method performing better than MA for raw spectral data.Compared to raw spectra,the R² values of the prediction sets for flowering and fruiting stages increased by 0.052 and 0.037,respectively,while RMSE decreased by 11.6%and 8.4%.In the flowering period model,the CNN model built using CARS-extracted feature wavelengths achieved superior prediction performance,with an R² of 0.815 and RMSE of 4.940 for the prediction set.In the fruiting period model,the RF model using UVE-CARS feature wavelengths achieved a prediction set R² of 0.875 and RMSE of 2.991,indicating high predictive ability.[Conclusion]By applying different preprocessing and wavelength extraction methods to hyperspectral data,and exploring machine learning-based models,we achieved rapid detection of leaf nitrogen concentration in greenhouse cucumbers at different growth periods,providing a theoretical basis for precise nitrogen fertilizer management.