Fine-resolution baseline maps of soil nutrients in farmland of Jiangxi Province using digital soil mapping and interpretable machine learning

• Published in: Catena (Giessen) 2025-02, Vol.249, Article 108635
• Main Authors: Hu, Bifeng, Geng, Yibo, Shi, Kejian, Xie, Modian, Ni, Hanjie, Zhu, Qian, Qiu, Yanru, Zhang, Yuan, Bourennane, Hocine
• Format: Article
• Language: English
• Subjects: Agricultural sciences; Digital soil mapping; Environmental Sciences; Fine maps; Life Sciences; Shapley additive explanations; Soil nutrients; Soil study; The main dominator
• ISSN: 0341-8162
• DOI: 10.1016/j.catena.2024.108635

[Display omitted] •Climate variables have dominant effects on mapping TN and TK.•Soil properties and climate variables made the largest contribution to map TP.•Introducing remote sensing images and soil management factors failed to improve prediction accuracy of soil nutrients.•Introducing CARS algorithm failed to improve prediction accuracy of soil nutrients.•Maps of location-specific primary covariate for different soil nutrients were produced. Detailed maps of soil nutrients are crucial for farmland management and agricultural production. However, soil nutrients are largely affected by various natural and anthropogenic factors, making it a challenging task to make clear its spatial distribution. To fill this gap, we produced the fine maps (30 m) of total content of nitrogen (TN), phosphorus (TP), and potassium (TK) in the farmland across Jiangxi Province in Southern China and quantified overall contribution of different covariates, as well as mapped the location-specific primary variable for predicting soil nutrients using an interpretable machine learning model. Our results reveal that random forest outperformed Cubist and XGBoost for mapping TN, TP and TK. The optimal models achieved R2 of 0.29, 0.29, 0.52 and RMSE of 0.43, 0.15 and 3.42 g kg−1 for TN, TP and TK, respectively. Moreover, we found both introducing competitive adaptive reweighted sampling algorithm and incorporating remote sensing images as well as soil management factors failed to clearly improve prediction accuracy of TN, TP and TK. In addition, climate variables had dominant overall effects on mapping TN (60.2 %) and TK (62.7 %), while soil properties made the largest contribution to mapping TP (34.3 %). The aridity index (46.90 %), mean annual solar radiation (34.94 %), and mean annual temperature (26.92 %) is the location-specific primary variable for mapping TN, TP, and TK in largest proportion of the study area, respectively. The soil nutrients maps we produced could function as baseline maps for monitoring spatio-temporal variation of soil nutrients, and our results could provide valuable implications for making more specific and efficient measures for soil management.