Active learning-based random forest algorithm used for soil texture classification mapping in Central Vietnam

[Display omitted] •Active learning model was explored for soil texture classification mapping.•Environmental covariate contribution level on soil texture classification was assessed.•Active learning model outperformed traditional model.•Soil textures with high sand content distributed mainly along the coast. In this study, an active learning model is introduced to map soil textures. We integrate density-based clustering (DBC) with a random forest (RF) algorithm to develop an active learning model. A dataset of 88 soil samples and eight environmental covariates, geology, altitude, slope, land use, distance to the sea, distance to waterways, distance to roads, and sediment transport index (STI), was used for training and testing at a scale of 70:30. Our results indicated that the traditional model achieved poor performance (accuracy = 53.3%) with weak agreement (kappa = 0.40), whereas, the active learning model achieved excellent performance (accuracy = 96.7%) with almost perfect agreement (kappa = 0.96). The five environmental covariates of distance to the sea, distance to waterways, distance to roads, altitude, and slope were the most important in explaining the soil texture classification. Approximately 70% of the total area was categorized as soil textures with high sand content, including sand, loamy sand, and sandy loam, distributed mainly along the coast. This work contributes a novel approach for soil texture classification mapping under limited budget, time, and human resource conditions.