MSRA-Net: A Multi-Task Learning Model for Soil Texture Prediction with Dynamic Weighting and Prior Knowledge Soft...

Area of Science:

• Agricultural Science
• Computer Science
• Soil Science

Background:

• Accurate soil texture data is vital for soil quality evaluation, conservation, and agricultural management.
• Convolutional Neural Networks (CNNs) offer superior accuracy in soil texture prediction compared to traditional machine learning.
• Existing lightweight models struggle with spectral modeling limitations like single-scale features and channel redundancy.

Purpose of the Study:

• To develop a novel, lightweight dynamic feature modeling approach for spectral data.
• To enhance feature representation and inter-channel interaction for improved spectral pattern capture.
• To introduce a multi-task learning variant for increased model stability and predictive accuracy.

Main Methods:

• Proposed the Multi-scale Routing Attention Network (MSRA-Net) integrating grouped multi-scale convolutions and intra-group Efficient Channel Attention (gECA).
• Implemented a Branch Routing Attention (BRA) mechanism for multi-scale weighting and enhanced feature interaction.
• Developed a multi-task learning variant (MSRA-MT) using uncertainty dynamic weighting to balance task gradients.

Main Results:

• MSRA-MT consistently outperformed baseline models on LUCAS and ICRAF datasets.
• Achieved strong performance and robustness with RMSEmean of 9.190 for ICRAF and 8.189 for LUCAS.
• Demonstrated that prior knowledge-based soft constraints can negatively impact optimization.

Conclusions:

• The proposed MSRA-Net and MSRA-MT offer effective solutions for lightweight spectral modeling in soil texture analysis.
• MSRA-MT shows significant improvements in predictive accuracy and robustness for soil texture mapping.
• Over-reliance on prior knowledge constraints may not always enhance learning effectiveness in soil science models.