Assessing the uncertainty of deep learning soil spectral models using Monte Carlo dropout

Abstract

The acquisition of soil information using infrared spectroscopy is now widely practised in soil sciences. In conjunction with machine learning models, spectral data are used to predict soil properties in an effort to complement or replace soil laboratory analyses. However, the assessment of uncertainty of spectral models is still not a common practice. Here, we evaluate two methods to assess the uncertainty of a deep learning soil spectral model for soil organic carbon prediction. We focused on how both methods evaluate in and out-of-domain uncertainty which signals if the model “knows” when it is making predictions on new data that differs from the data used during training, where it is expected to generate wider prediction intervals. The methods corresponded to a) bootstrapping, a general framework for uncertainty assessment commonly used in soil modelling and b) Monte Carlo dropout, specifically used to assess the uncertainty of deep learning models. Using the LUCAS Vis–NIR spectra database for predicting soil organic carbon, we demonstrate that both models correctly assessed the uncertainty for in-domain data, showing similar results. However, when predicting on out-of-domain data, Monte Carlo dropout was capable of generating much wider prediction intervals compared to bootstrapping, correctly signalling that the new data is different from the data used during training. This uncertainty should accompany the prediction and be propagated when the prediction is being used in downstream applications.