Abstract. Digital soil mapping (DSM) has been widely used as a cost-effectivemethod for generating soil maps. However, current DSM data representationrarely incorporates contextual information of the landscape. DSM models areusually calibrated using point observations intersected with spatiallycorresponding point covariates. Here, we demonstrate the use of theconvolutional neural network (CNN) model that incorporates contextual informationsurrounding an observation to significantly improve the prediction accuracyover conventional DSM models. We describe a CNN model that takes inputs as images of covariates and explores spatialcontextual information by finding non-linear local spatial relationships ofneighbouring pixels. Unique features of the proposed model include inputrepresented as a 3-D stack of images, data augmentation to reduce overfitting,and the simultaneous prediction of multiple outputs. Using a soil mapping examplein Chile, the CNN model was trained to simultaneously predict soil organiccarbon at multiples depths across the country. The results showed that, inthis study, the CNN model reduced the error by 30% compared withconventional techniques that only used point information of covariates. Inthe example of country-wide mapping at 100m resolution, the neighbourhoodsize from 3?to 9?pixels is more effective than at a point location and largerneighbourhood sizes. In addition, the CNN model produces less predictionuncertainty and it is able to predict soil carbon at deeper soil layers moreaccurately. Because the CNN model takes the covariate represented as images, itoffers a simple and effective framework for future DSM models.
展开▼
