Inversion model of soil salinity in alfalfa covered farmland based on sensitive variable selection and machine...

Area of Science:

• Agricultural remote sensing
• Soil science
• Machine learning applications

Background:

• Accurate soil salinity content (SSC) monitoring is vital for efficient irrigation management in large-scale agriculture.
• Uncrewed aerial vehicle (UAV) low-altitude remote sensing offers high spatial and temporal resolution for SSC monitoring.
• Existing models often lack comprehensive evaluation of variable selection methods combined with machine learning algorithms.

Purpose of the Study:

• To investigate the effectiveness of combining different variable selection methods with machine learning algorithms for soil salinity inversion.
• To identify the optimal combination of variable selection and machine learning for accurate SSC monitoring in crop-covered farmland.
• To evaluate the performance of various models using metrics like R², RMSE, and RPD.

Main Methods:

• Extracted spectral indices from UAV multispectral data.
• Employed four variable selection methods: Pearson correlation coefficient (PCC), gray relational analysis (GRA), variable projection importance (VIP), and support vector machine-recursive feature elimination (SVM-RFE).
• Developed 20 soil salinity inversion models using Support Vector Machine regression (SVM), Back Propagation Neural Network (BPNN), Extreme Learning Machine (ELM), and Random Forest (RF) algorithms, comparing screened and unscreened variables.

Main Results:

• Variable selection combined with machine learning significantly enhanced soil salinity inversion accuracy.
• Gray relational analysis (GRA) was suitable for SVM, BPNN, and ELM, while PCC was best for RF.
• The GRA-SVM model achieved the highest accuracy (Rv²=0.8888, RMSEv=0.1780, RPD=1.8115) for alfalfa-covered farmland.

Conclusions:

• Combining variable selection methods with machine learning algorithms is a highly effective approach for improving remote sensing-based soil salinity inversion.
• This study provides a robust methodology for accurate and timely SSC information acquisition in crop-covered farmlands.
• The GRA-SVM model offers a reliable solution for soil salinity mapping and management in arid oasis irrigation areas.