Accuracy and uncertainty of digital soil mapping approaches to extract and transfer soil information from reference area

• Autor(a) principal: Machado, Diego Fernandes Terra
• Data de Publicação: 2017
• Tipo de documento: Dissertação
• Idioma: eng
• Título da fonte: Repositório Institucional da UFLA
• Texto Completo: http://repositorio.ufla.br/jspui/handle/1/28220
• topic: Mapeamento do solo – Modelos; Soil mapping – Models; Ciência do Solo

Contradicting the need for detailed maps, we currently experience scarcity of investments on soil surveys in Brazil. In this sense, it is necessary to resort to techniques that allow the expansion of the mapped areas, at relatively lower costs. From this perspective, this work focused on the investigation of procedures and tools for the retrieving and extrapolation of soil type information from a reference area to its surroundings. The objectives included: (i) retrieving information from a detailed soil map of a reference area; (ii) to evaluate the transferability of information to a larger area, which preserves similar environmental characteristics similar to those of the pilot area; (iii) evaluate the accuracy and uncertainty of the inference models. From a Digital Elevation Model, a series of topographic indexes were calculated, which were correlated with the soil classes, represented by mapping units of the legacy map. The objective was to infer from the soil-landscape relationship of the pilot area, the distribution of soil types in the extrapolation area. For that duty, three inference procedures were applied, one data-driven (Random Forest (RF)) and two others, based on knowledge (Rule-based reasoning and Case-based reasoning - ArcSIE). Regarding RF, 52 models were graded from a routine of tuning and different combinations of training data. Although considered a robust predictor, RF demonstrated sensitivity to training strategies. Most of the models presented low accuracy. However, at least one model with more than 80% of global accuracy was obtained. Regarding RBR and CBR procedures, only the former resulted in a map with good precision. The advantage of using knowledge-based systems like RBR is to make explicit the soil-landscape relationship through a systematic set of rules. By accessing the uncertainty of the predictions, in addition to evaluating the behavior of the models, it was possible to observe the complexity of the soil-landscape relationship of Oxisols and Inceptisols, characteristic of tropical environments. This is particularly important for model review and sampling planning in the search for more accurate maps.