Simultaneous Prediction of Soil Properties Using Multi_CNN Model.

Area of Science:

• Agricultural Science
• Spectroscopy
• Machine Learning

Background:

• Near-infrared spectroscopy (NIRS) is crucial for rapid soil information acquisition.
• Deep learning significantly improves traditional soil prediction models.
• Current soil prediction models often suffer from low efficiency and accuracy.

Purpose of the Study:

• To propose an intelligent soil multi-attribute prediction method using deep learning.
• To develop a dual-stream convolutional neural network (CNN) model for enhanced soil attribute prediction.
• To achieve simultaneous prediction of multiple soil attributes through feature fusion.

Main Methods:

• Constructed a dual-stream CNN model (Multi_CNN) integrating 1D and 2D convolutions.
• Extracted soil attribute features from both spectral sequences and spectrograms.
• Validated the model on two datasets of varying scales for multi-attribute prediction.

Main Results:

• Achieved high prediction accuracy (RP2) for Total Carbon (0.94), Total Nitrogen (0.95), and Alkaline Nitrogen (0.87) on a small dataset.
• Obtained strong RP2 values for Organic Carbon (0.95), Nitrogen (0.91), and Clay (0.83) on the LUCAS dataset.
• Demonstrated superior accuracy compared to traditional regression and other contemporary soil nutrient prediction methods.

Conclusions:

• The proposed Multi_CNN model offers a more accurate and efficient approach to soil multi-attribute intelligent prediction.
• Combining spectral sequence and spectrogram analysis in a dual-stream CNN architecture is effective for NIRS-based soil analysis.
• This deep learning method advances the field of rapid soil information acquisition and analysis.