Predictive soil mapping based on the similarity of environmental covariates using a spatial convolutional autoencoder

Abstract

AbstractIndividual predictive soil mapping (iPSM) can predict the soil properties and quantify the prediction uncertainty by using a limited quantity of soil samples. This method assumes that the locations with similar environmental conditions have similar soil properties. The similarity between the locations in the iPSM is calculated based on the environment covariates corresponding to each location. However, this method does not consider the spatial structure information of covariates at the locations. To address this shortcoming, this study proposes a convolutional autoencoder (CAE) model to obtain new covariates called the CAE covariates. These covariates represent the spatial structure information of environment covariates at a certain location. The iPSM based on the similarity calculated using the CAE covariates is applied to predict clay, silt, and sand. A total of 60 sets of compressed CAE covariates are generated based on different spatial neighborhoods and embedded layer sizes. Compared to the original covariates, the iPSM based on the similarity calculated by the CAE covariates improves the prediction accuracy and significantly reduces the prediction uncertainty for clay, silt, and sand. For embedded 20 of neighborhood 9, the improvements in clay, silt, and sand are the highest, and they are 20.2%, 15.8%, and 18.5%, respectively. The largest improvement in the average uncertainty is 96.4% for embedded 5 of neighborhood 13. By applying a range of user‐defined thresholds, the proposed method improves the prediction accuracy for clay, silt, and sand, reduces the prediction uncertainty, and expands the prediction area simultaneously. The trade‐off between the prediction accuracy and the areas that can be predicted by the iPSM is alleviated. The CAE covariates are promising for other machine learning algorithms for digital soil mapping and models that require using spatial environment variables as an input.