Identifying optimized spectral and spatial features of UAV-based RGB and multispectral images to improve potato nitrogen content estimation

作者：Hang Yin, Haibo Yang, Yuncai Hu, Fei Li, Kang Yu · 发表于：Smart Agricultural Technology · 年份：2025 · DOI：10.1016/j.atech.2025.101255 · 被引用次数：3 · 研究领域：Remote Sensing in Agriculture、Spectroscopy and Chemometric Analyses、Smart Agriculture and AI

• Determine the optimal input variables for machine learning models. • RGB-based texture and MS-based spectral indices are the most important for PNC prediction. • The combination of RGB-based texture and MS-based spectral indices with VHGPR produces the highest prediction accuracy. Selecting optimal input features for the machine learning (ML) models is crucial in estimating plant nitrogen content (PNC). The goals of this study were to (i) identify optimal spectral indices and texture features from RGB and multispectral (MS) images and (ii) improve the accuracy of PNC prediction by combining optimal features with ML. The standard and variational heteroscedastic Gaussian process regression (SGPR and VHGPR) were trained using different image features from the 2020-2021 potato nitrogen fertilizer experiments. Based on the evaluation of RReliefF algorithm, the RGB-based multi-scale texture and MS-based spectral indices were the most important for PNC. The root mean square error (RMSE) of the models using RGB-based texture (RMSE = 0.36-0.39%) was slightly lower than those based on spectral indices (RMSE = 0.38-0.40%). For the combined RGB and MS data, spectral indices were the important input features for the SGPR and VHGPR models. The optimized mND705 was selected as the most critical spectral indices for PNC prediction. Incorporating RGB-based texture features and MS-based spectral indices into the VHGPR model achieved the highest prediction accuracy (RMSE = 0.29%). In conclusi...