A MSCA-CGC model for simultaneously predicting multiple soil properties via Vis–NIR spectral data

In the simultaneous prediction of soil properties, visible–near infrared (Vis–NIR) spectroscopy is widely used for rapid and non-destructive assessment of soil properties. However, traditional multi-task learning (MTL) approaches in such multivariate prediction tasks often show strong dependency on task correlation, limiting their robustness and generalization. To address this limitation, this study proposes the MSCA-CGC, a network that utilises Multi-Scale Convolution-Channel Attention within a Customised Gate Control framework. The model combines shared and task-specific experts to better capture local high-frequency spectral features. Each task uses an independent gating network to dynamically fuse shared and task-specific information, reducing the seesaw effect and negative transfer. Using the LUCAS 2009 dataset (19,032 samples, 11 soil properties), the model is benchmarked against PLSR, XGBoost, MMoE, and LSTM-CNN-Attention. Results show that MSCA-CGC outperforms the other models in prediction accuracy and stability, with average RMSE reduced by 3%–33%, R2 improved by 3%–39%, and RPD increased by 3%–51%. In transfer tests on LUCAS 2015, MSCA-CGC demonstrated good generalisation for key properties, supporting its robustness and adaptability under spatiotemporal heterogeneity. This study confirms the effectiveness of MSCA-CGC for soil property prediction.