Digital terrain models (DTM) have been used in soil mapping worldwide. When using such models, improved predictions are often attained with the input of extra variables provided by the use of proximal sensors, such as magnetometers and portable X-ray fluorescence scanners (pXRF). This work aimed to evaluate the efficiency of such tools for mapping soil classes and properties in tropical conditions. Soils were classified and sampled at 39 locations in a regular-grid design with a 200-m distance between samples. A pXRF and a magnetometer were used in all samples, and DTM values were obtained for every sampling site. Through visual analysis, boxplots were used to identify the best variables for distinguishing soil classes, which were further mapped using fuzzy logic. The map was then validated in the field. An ordinary least square regression model was used to predict sand and clay contents using DTM, pXRF and the magnetometer as predicting variables. Variables obtained with pXRF showed a greater ability for predicting soil classes (overall accuracy of 78% and 0.67 kappa index), as well as for estimating sand and clay contents than those acquired with DTM and the magnetometer. This study showed that pXRF offers additional variables that are key for mapping soils and predicting soil properties at a detailed scale. This would not be possible using only DTM or magnetic susceptibility.

The permanent monitoring of vegetation cover is important to guarantee a sustainable management of agricultural activities, with a relevant role in the reduction of water erosion. This monitoring can be carried out through different indicators such as vegetation cover indices. In this study, the vegetation cover index was obtained using uncalibrated RGB images generated from a digital photographic camera on an unmanned aerial vehicle (UAV). In addition, a comparative study with 11 vegetation indices was carried out. The vegetation indices CIVE and EXG presented a better performance and the index WI presented the worst performance in the vegetation classification during the cycles of jack bean and millet, according to the overall accuracy and Kappa coefficient. Vegetation indices were effective tools in obtaining soil cover index when compared to the standard Stocking method, except for the index WI. Architecture and cycle of millet and jack bean influenced the behavior of the studied vegetation indices. Vegetation indices generated from RGB images obtained by UAV were more practical and efficient, allowing a more frequent monitoring and in a wider area during the crop cycle.

Different uses of soil legacy data such as training dataset as well as the selection of soil environmental covariables could drive the accuracy of machine learning techniques. Thus, this study evaluated the ability of the Random Forest algorithm to predict soil classes from different training datasets and extrapolate such information to a similar area. The following training datasets were extracted from legacy data: a) point data composed of 53 soil samples; b) 30 m buffer around the soil samples, and soil map polygons excluding: c) 20 m; and d) 30 m from the boundaries of polygons. These four datasets were submitted to principal component analysis (PCA) to reduce multidimensionality. Each dataset derived a new one. Different combinations of predictor variables were tested. A total of 52 models were evaluated by means of error of models, prediction uncertainty and external validation for overall accuracy and Kappa index. The best result was obtained by reducing the number of predictors with the PCA along with information from the buffer around the points. Although Random Forest has been considered a robust spatial predictor model, it was clear it is sensitive to different strategies of selecting training dataset. Effort was necessary to find the best training dataset for achieving a suitable level of accuracy of spatial prediction. To identify a specific dataset seems to be better than using a great number of variables or a large volume of training data. The efforts made allowed for the accurate acquisition of a mapped area 15.5 times larger than the reference area.

The estimation of soil physical and chemical properties at non-sampled areas is valuable information for land management, sustainability and water yield. This work aimed to model and map soil physical-chemical properties by means of knowledge-based digital soil mapping approach as a study case in two watersheds representative of different physiographical regions in Brazil. Two watersheds with contrasting soil-landscape features were studied regarding the spatial modeling and prediction of physical and chemical properties. Since the method uses only one value of soil property for each soil type, the way of choosing typical values as well the role of land use as a covariate in the prediction were tested. Mean prediction error (MPE) and root mean square prediction error (RMSPE) were used to assess the accuracy of the prediction methods. The knowledge-based digital soil mapping by means of fuzzy logics is an accurate option for spatial prediction of soil properties considering: 1) lesser intense sampling scheme; 2) scarce financial resources for intensive sampling in Brazil; 3) adequacy to properties with non-linearity distribution, such as saturated hydraulic conductivity. Land use seems to influence spatial distribution of soil properties thus, it was applied in the soil modeling and prediction. The way of choosing typical values for each condition varied not only according to the prediction method, but also with the nature of spatial distribution of each soil property.