Enhanced Prediction of Soil Carbon via Encoder-Decoder Neural Networks for a Boreal Study Area in Northern Ontario

Abstract

Addressing the impacts of carbon in connection with land cover conversion and climate change is of predominant interest for boreal realms. Consequently, boosting accuracy for the prediction of total carbon (C) with soil mapping is a crucial objective, particularly for a boreal study area under risk of land cover transition in northern Ontario, Canada. To enhance the prediction of soil modeling, integrated approaches combining encoder-decoder (ED) with dense neural network (DNN) and convolutional neural network (CNN) formulations suitable for smaller target data sets were developed. These methods were able to effectively extract dominant features within predictor data and augment modeling accuracy. The obtained results were compared with those attained from structural equation modeling (SEM) and random forest (RF), as well as basic DNN and CNN models. A model ensemble based on all approaches was also considered, from which standard deviations were calculated to gauge the prediction uncertainty. Quantile mappings with respect to deciles were also derived from the model ensemble to provide additional insights with prediction. Validation accuracies for the ED-CNN model attained a coefficient of determination (R2) of 0.59. The greatest deviations with predicting C contents corresponded to the wetlands. However, when quantified by decile mapping, forested localities within river valleys encountered the highest uncertainties with prediction, indicting a need for better modeling of sites with intermediate concentrations of soil C.