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

## TL;DR

MSRA-Net is a new lightweight model for predicting soil texture using improved neural network techniques and multi-task learning, achieving better accuracy and robustness.

## Contribution

The novel MSRA-Net model introduces dynamic weighting and multi-scale routing attention for improved soil texture prediction with lightweight CNNs.

## Key findings

- MSRA-MT outperforms baseline models on LUCAS and ICRAF datasets with RMSEmean of 9.190 and 8.189, respectively.
- Dynamic weighting in multi-task learning improves model stability and predictive accuracy.
- Prior knowledge soft constraints may amplify noise and hinder optimization.

## Abstract

Accurate and rapid acquisition of soil texture information is crucial to evaluating soil quality, formulating soil and water conservation strategies, and guiding agricultural resource management. Compared with traditional machine learning methods, convolutional neural networks (CNNs) demonstrate superior accuracy in soil texture prediction. To overcome the limitations of existing lightweight models in spectral modeling, such as insufficient single-scale feature representation, limited channel utilization, and branch redundancy, and to meet the demand for lightweight architectures, we propose a novel dynamic feature modeling approach: Multi-scale Routing Attention Network (MSRA-Net). MSRA-Net integrates grouped multi-scale convolutions with an intra-group Efficient Channel Attention (gECA) mechanism, combined with a multi-scale weighting strategy based on a Branch Routing Attention (BRA) mechanism, thereby enhancing inter-channel feature interaction and improving the model’s ability to capture complex spectral patterns. Furthermore, we introduce a multi-task learning variant, MSRA-MT, which employs uncertainty dynamic weighting to balance gradients magnitude across tasks, thereby improving both stability and predictive accuracy. Experimental results on the LUCAS and ICRAF datasets demonstrate that the MSRA-MT model consistently outperforms baseline models in terms of performance and robustness (RMSEmean = 9.190 and RMSEmean = 8.189 for ICRAF and LUCAS, respectively). Prior knowledge-based soft constraints may hinder optimization by amplifying intrinsic noise, rather than improving learning effectiveness.