Inversion of Soil Organic Matter Content Based on Improved Convolutional Neural Network

Area of Science:

• Agricultural Science
• Remote Sensing
• Data Science

Background:

• Soil organic matter (SOM) is crucial for crop nutrition and soil health.
• Accurate SOM prediction is vital for sustainable agriculture and land management.
• Traditional methods for SOM assessment can be time-consuming and resource-intensive.

Purpose of the Study:

• To investigate the feasibility of using deep learning for multi-feature soil organic matter (SOM) prediction.
• To develop and optimize a deep learning model for SOM content inversion using remote sensing data.
• To evaluate the impact of feature selection on model performance and computational efficiency.

Main Methods:

• Utilized Sentinel-2A remote sensing images from Nong'an County for dataset construction.
• Employed field sampling and image processing techniques to gather and prepare data.
• Adapted and improved the LeNet-5 convolutional neural network for SOM prediction.
• Compared the deep learning model's performance against linear regression, Support Vector Machines (SVM), and Random Forest (RF) using RMSE and R² metrics.

Main Results:

• The improved deep learning model achieved a high coefficient of determination (R²) of 0.89 and a minimum RMSE of 2.979 for SOM prediction.
• The deep learning model significantly outperformed SVM (R² = 0.4035) and RF (R² = 0.4913).
• Feature importance analysis revealed that selecting optimal features improved prediction results and enhanced computational speed compared to using all features.

Conclusions:

• Deep learning models are effective for inverting soil organic matter content using multispectral remote sensing data.
• Optimizing feature selection is critical for improving the accuracy and efficiency of SOM prediction models.
• This approach offers a promising, data-driven method for large-scale SOM assessment in agriculture.