Modelling the extent of northern peat soil and its uncertainty with Sentinel: Scotland as example of highly cloudy region

Abstract Mapping the extent and locations of peatland at landscape scale has implications for carbon inventories, conservation and ecosystem services assessments. The main aim of this paper was to model and map the extent of northern peat soils while taking into account its uncertainty, and in particular exploring: 1. the use of radar Sentinel 1 as alternative to optical sensors to reduce problems due to clouds while taking advantage of the seasonality changes and 2. the use of deep learning for peat classification and as application of Digital Soil Mapping. The data sets defining presence or absence of peat in the soil were obtained from different sources and different sampling schemes, densities and distributions. Scotland was used as test case, because of its cloudy weather and fragmented distribution of different types of peat. An extension of the scorpan-kriging approach was used. The trend was estimated with different approaches: Generalized Additive Models, RandomForest and deep learning (convolutional neural networks). Each approach produced the probability of belonging to a class and the predicted class for each pixel. The results were assessed using out-of-sample measures. In this study 108 combinations of data sets and models (including trend approaches, sets of covariates and modelling of the spatial structure) were assessed. Overall, spatially explicit models performed better. The choice of the statistical method can have a significant impact on the predictive performances, while the sets of environmental covariates had a lower impact. Sentinel-1 with morphological features proved to be a good alternative to optical data for peat mapping. It is important to have balanced data sets representing the distribution of the data, because merging heterogeneous sources of data from different populations does not necessarily improve predictions. The use of deep learning and convolutional neural network provided initial promising results. There were large differences in the modelling approaches. These differences and uncertainties need to be taken into account for further modelling such as earth surface modelling or carbon accounting.