Interpretation of Convolutional Neural Networks for Acid Sulfate Soil Classification

Amélie Beucher,Mogens H Greve,Christoffer B Rasmussen,Thomas B Moeslund

doi:10.3389/fenvs.2021.809995

Abstract

Convolutional neural networks (CNNs) have been originally used for computer vision tasks, such as image classification. While several digital soil mapping studies have been assessing these deep learning algorithms for the prediction of soil properties, their potential for soil classification has not been explored yet. Moreover, the use of deep learning and neural networks in general has often raised concerns because of their presumed low interpretability (i.e., the black box pitfall). However, a recent and fast-developing sub-field of Artificial Intelligence (AI) called explainable AI (XAI) aims to clarify complex models such as CNNs in a systematic and interpretable manner. For example, it is possible to apply model-agnostic interpretation methods to extract interpretations from any machine learning model. In particular, SHAP (SHapley Additive exPlanations) is a method to explain individual predictions: SHAP values represent the contribution of a covariate to the final model predictions. The present study aimed at, first, evaluating the use of CNNs for the classification of potential acid sulfate soils located in the wetland areas of Jutland, Denmark (c. 6,500 km2), and second and most importantly, applying a model-agnostic interpretation method on the resulting CNN model. About 5,900 soil observations and 14 environmental covariates, including a digital elevation model and derived terrain attributes, were utilized as input data. The selected CNN model yielded slightly higher prediction accuracy than the random forest models which were using original or scaled covariates. These results can be explained by the use of a common variable selection method, namely recursive feature elimination, which was based on random forest and thus optimized the selection for this method. Notably, the SHAP method results enabled to clarify the CNN model predictions, in particular through the spatial interpretation of the most important covariates, which constitutes a crucial development for digital soil mapping.

Highlights

Digital soil mapping (DSM) techniques traditionally relate soil observations with point information extracted from different environmental covariates at corresponding locations
During the last 2 decades, various studies attempted to incorporate spatial context through the preprocessing of environmental covariates. These original contextual mapping studies have used: different neighboring size to compute terrain attributes (Smith et al, 2006; Behrens et al, 2010b), adaptative filters applied on elevation data or derived attributes (Behrens et al, 2018b), covariates spatially transformed with wavelet analysis (Lark and Webster, 2001; Sun et al, 2017), hypercovariates created from elevation data (Behrens et al, 2010a; Behrens et al, 2014), and a multi-scale approach using aggregated covariates (Miller et al, 2015)
RForiginal and RFscaled models yielded comparable overall prediction accuracy (OA) values for validation, 61 and 63%, respectively. These results indicate that the RFscaled model did not seem to use the contextual spatial information to classify potential acid sulfate (AS) soils more accurately than the RForiginal model

Summary

Introduction

Digital soil mapping (DSM) techniques traditionally relate soil observations with point information extracted from different environmental covariates at corresponding locations. During the last 2 decades, various studies attempted to incorporate spatial context through the preprocessing of environmental covariates These original contextual mapping studies have used: different neighboring size to compute terrain attributes (Smith et al, 2006; Behrens et al, 2010b), adaptative filters applied on elevation data or derived attributes (Behrens et al, 2018b), covariates spatially transformed with wavelet analysis (Lark and Webster, 2001; Sun et al, 2017), hypercovariates created from elevation data (Behrens et al, 2010a; Behrens et al, 2014), and a multi-scale approach using aggregated covariates (Miller et al, 2015). Other DSM studies investigated the use of DNNs for mapping soil moisture content (Song et al, 2016) and soil organic carbon content (Taghizadeh-Mehrjardi et al, 2020b; Emadi et al, 2020; Tao et al, 2020), identifying and delineating soil master horizons (Jiang et al, 2021), or spectral modelling (Singh and Kasana, 2019; Gholizadeh et al, 2020)

Methods

Results

Conclusion