Towards the Improvement of Soil Salinity Mapping in a Data-Scarce Context Using Sentinel-2 Images in Machine-Learning...

Area of Science:

• Remote Sensing
• Soil Science
• Machine Learning

Background:

• Soil salinity monitoring is crucial but challenging due to costly and time-consuming electrical conductivity (EC) measurements.
• Existing machine learning models for soil salinity mapping often suffer from limited training data, impacting consistency.
• Sentinel-2 reflectance data offers a viable input for soil salinity mapping.

Purpose of the Study:

• To enhance machine learning models for soil salinity mapping in data-scarce environments.
• To evaluate a novel method for expanding limited soil salinity datasets.
• To compare the performance of Random Forest (RF) and Support Vector Machine (SVM) models using enhanced datasets.

Main Methods:

• A method was developed to expand the original dataset (OD) by assigning EC values from sampled pixels to their eight neighbors, creating an enhanced dataset (ED).
• Two machine learning models, Random Forest (RF) and Support Vector Machine (SVM), were trained using both OD and ED.
• Model performance was assessed by comparing predictions with independent EC observations, and feature selection techniques (VIF, GA) were applied.

Main Results:

• The enhanced dataset (ED) significantly improved model consistency, with overall accuracy increasing from 0.25 (0.26) for OD to 0.77 (0.55) for ED in RF (SVM) models.
• The Random Forest model demonstrated superior performance in soil salinity estimation compared to the Support Vector Machine model.
• Feature selection, particularly using the Genetic Algorithm (GA), further enhanced model reliability.

Conclusions:

• The proposed method of expanding the learning database is effective for improving soil salinity mapping accuracy in data-scarce conditions.
• Combining machine learning with Sentinel-2 data holds significant potential for soil salinity monitoring.
• Optimal machine learning setup requires careful consideration of both model choice and feature selection methods.