Spatial heterogeneity response of soil salinization inversion cotton field expansion based on deep learning

Area of Science:

• Environmental Science
• Agronomy
• Remote Sensing

Background:

• Soil salinization is a major environmental issue in arid regions, impacting food security and requiring effective management strategies.
• Existing machine learning models struggle to capture localized salinity data and the spatial variability of salinity's effects on crops.

Purpose of the Study:

• To develop and compare Convolutional Neural Network (CNN), Long Short-Term Memory Network (LSTM), and Random Forest (RF) models for soil salinization inversion.
• To accurately map soil salinity and analyze its relationship with cotton field expansion.
• To identify areas vulnerable to soil salinity for targeted management.

Main Methods:

• Utilized 97 field samples and Landsat-8 imagery for feature extraction.
• Developed and evaluated CNN, LSTM, and RF models for soil salinization prediction.
• Selected the best-performing model (CNN) for high-resolution (30m) soil salinity mapping in 2013 and 2022.

Main Results:

• The CNN model demonstrated superior performance (R²=0.84 training, R²=0.73 test) in capturing local salinity information.
• Cotton field expansion significantly decreased soil salinization, reducing severely salinized and saline soil areas in new fields.
• Areas with long-term cotton cultivation and newly reclaimed fields show high sensitivity and vulnerability to soil salinity.

Conclusions:

• Deep learning models, particularly CNN, offer excellent performance for precise soil salinization mapping.
• Cotton cultivation can mitigate soil salinization, contributing to sustainable agriculture and food security.
• Understanding spatial correlations between soil salinity and crop distribution is crucial for effective land and salinity management.