Using the farmers’ knowledge on soils improves local digital soil mapping products: A case study in South India

• We explored how local farmers’ soil knowledge could provide new input for DSM. • Farmers’ soil types and their location were collected in two participatory workshops. • Farmers’ soil types localisations and the soil map were in fairly good agreement. • Adding such locations to DSM modelling significantly improved soil class predictions. • Coupling DSM with ethnopedology increases soil mapping accuracy and credibility. As ethnopedologists have observed, farmers have accumulated valuable knowledge of local spatial variations in soils, which they are currently mobilising for their own use. Such knowledge could complement the soil databases that provide the usual soil inputs of Digital Soil Mapping (DSM) approaches. This paper presents a novel participatory DSM approach that uses the collective knowledge of farmers living in a small area (Gopalapura village, South India) as input to DSM models that predict soil texture classes, depth and stoniness. The collection of farmers’ soil knowledge involved two steps. First, a local soil typology was created by a group of 27 farmers. Secondly, the farmers were asked to locate each soil type within the area of interest. The output was a set of locations labelled with a soil type, which was used as additional input for training a random forest in two ways: i) as an alternative soil input dataset, and ii) as a new set of covariates by calculating topography-oriented buffer distances from the farmers’ soil types. In order to evaluate these new soil data and their potential contribution to improving the performance of the DSM, the localisations of farmer soil types, as well as the predicted maps of texture, depth and stoniness classes, were compared with a very detailed (1:8,000) soil map. The results showed fairly good agreement of soil class determinations between the soil map and the farmer’s soil information. However, using the farmer soil types as an alternative soil input to train a random forest gave poor prediction performances. Conversely, adding topography-oriented distances to farmer soil types to the usual set of covariates significantly improved the predictions regardless of the soil property tested and the density of soil sampling. The hybridisation of digital soil mapping and ethnopedological approaches seems promising both for improving the quality of local soil maps and for ensuring a better appropriation of such maps by future local users.