PS-MTL-LUCAS: A partially shared multi-task learning model for simultaneously predicting multiple soil properties
Zhaoyu Zhai, Chen Fuji, Hongfeng Yu, Simon Hu, Xinfei Zhou, Huanliang Xu
Abstract
Soil acts as a foundation for human survival and social development and soil quality has a great effect on the growth of agricultural products. Visible/near-infrared spectroscopy has been acknowledged as a rapid and non-destructive method for predicting soil properties, and multi-task learning is a preferable approach to analyze the correlation between the spectroscopy data and soil properties. However, current multi-task learning models with the soft parameter sharing structure extremely rely on the task relatedness. To tackle this limitation, we proposed PS-MTL-LUCAS, a multi-task learning with a partially shared structure in this study. An additional shared layer was utilized to learn the general informative representations and interact with each task-specific layer. The partially shared structure ensured the maximum information flow between layers, thereby boosting the prediction performance. Also, the SHapley Addictive exPlanations (SHAP) algorithm was adopted to extract the feature wavelengths of each soil property. PS-MTL-LUCAS was validated on the LUCAS topsoil dataset (2009), and the experimental result suggested that PS-MTL-LUCAS dominated state-of-the-art models by achieving the determination of coefficient at 0.945, 0.936, 0.413, 0.624, 0.837, 0.952, and 0.956 for pH, N, P, K, CEC, OC, and CaCO3, respectively. In summary, this study highlighted the use of the soil spectroscopy and multi-task learning techniques in the soil property prediction task and provided a very promising approach for soil studies.