.png){kind=icon}
.png){kind=icon}
Return
Soil salinity inversion based on UAV hyperspectral data and stacking ensemble learning
.png){kind=icon}
PRE
AI
T
Q
H
Z
Z
X
DOI：10.1080/01431161.2026.2693200.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Soil salinization severely constrains sustainable agricultural development. Accurate monitoring of soil salinity in farmland is therefore of great significance for saline – alkali land management and food security. In this study, typical saline farmland in Caofeidian District, Tangshan, Hebei, China, was selected as the study area, and three experimental plots with different salinization levels were investigated. Considering that the spectral response of soil salinity is relatively weak and that traditional multispectral satellite remote sensing data have limited capability to finely characterize the spatial heterogeneity of salinity, a UAV hyperspectral soil salinity inversion model based on a stacking ensemble framework was developed. Spectral bands were first preprocessed using a combination of logarithmic transformation and standard normal variate transformation. Subsequently, a two-stage feature selection strategy, CARS – SHAP, was proposed to identify informative spectral bands. Specifically, the Competitive Adaptive Reweighted Sampling (CARS) method was first used to select 50 major spectral bands based on regression coefficient weights to reduce redundancy, and multi-model SHAP analysis was then applied to identify 10 key salinity-sensitive spectral bands with stable contributions across different models, thereby improving feature selection stability and prediction consistency under different salinization conditions. Finally, a soil salinity inversion model was constructed using CatBoost, XGBoost, and a multilayer perceptron (MLP) as base learners and partial least squares regression (PLSR) as the meta-learner within the stacking framework. Experimental results demonstrate that the proposed model outperforms conventional single models in both prediction accuracy and stability, achieving an R2 of 0.925 and an RMSE of 0.560 g kg−1 on the validation set. The results provide technical support for high-precision remote sensing monitoring of farmland soil salinity and precision agricultural management.
Keywords:
UAV hyperspectral data
soil salinity inversion
ensemble learning
band selection
stacking
Journal
IF:
2.6
Papers:
1.2W
Citations:
2.7W
Organization
Cited Papers