A novel local-global dependency deep learning model for soil mapping

Abstract

The accurate and cost-effective mapping of soil texture is essential for agricultural development and environmental activities. Soil texture exhibits high spatial heterogeneity which poses challenges for recent Digital Soil Mapping (DSM) methods in achieving accurate predictions. Feature engineering methods, extensively used to capture complex soil-forming relationships and enhance prediction accuracy, often involve labor-intensive processes. Additionally, the engineered “discrete” feature cannot reflect interactions between environmental covariates or dependencies. To address the challenges, this study proposes a novel Local-Global Dependency Long Short-Term Memory model (LGD-LSTM) to enhance soil texture predictions at various soil depths. Firstly,a covariate reorganization method hasbeen devised to generate multiple sets of input. Subsequently, several Long Short-Term Memory models (LSTM)have beenemployed to extract the interdependencies among the covariates.Finally, predictions are generated using a fully-connected layer. Cross-validation was conducted within this experiment to analyze prediction accuracy: the average explained variation (R2) ranged from 0.66 to 0.73, and the root mean square error (RMSE) ranged from 6.52% to 10.89%. The results indicated that the LGD-LSTM model offers distinct advantages over other digital soil mapping methods, including Random Forests (RF), Convolutional Neural Network (CNN), and the standard Long Short-Term Memory model (LSTM). In summary, this LGD-LSTM method demonstrates superior performance with relatively high accuracy, ensuring its applicability in effectively representing spatial variations in soil texture. Furthermore, it presents a novel option for DSM applications, enhancing the field's methodology and potential impact.