Mapeamento digital de atributos do solo em áreas remotas sob Floresta Amazônica: um estudo de caso na formação Solimões

Abstract

Currently, the demand for digital soil mapping (DSM) has grown considerably in environmental studies. Faced with these demands and the lack of information on soils on an adequate scale, it is more than necessary to develop research, techniques and methods that help in the study of soils and their attributes. This study was divided into two chapters: The first chapter aimed to evaluate the application of machine learning models (ML) in the prediction of soil carbon stock at 30cm (SOCS30) and 100cm (SOCS100) depth and in the second chapter in the prediction in surface and subsurface the soil texture composition. The specific objectives were; Compare two forms of dataset distribution; reference area (RA) and total area (TA), to evaluate two categories of covariate selection: "wrapper method" and "pre-selection of covariate" as a pre-processing step, before ML calibration, and to evaluate the performance of three ML algorithms: regression tree (RT), random forest (RF) and support vector machine (SVM) in the modeling of soil attributes. The study site was divided into three blocks, called Urucu, Araracanga and Juruá blocks. The dataset consisted of 120 carbon stock observations and 151 particle size

composition observations combined with 21 covariates (20 relief covariates and 1 P-band- derived index from radar) that were addressed in the two different datasets: (RA) and (TA). The

similarity between the RA and the landscape of the blocks was evaluated using the Gower general index and the descriptive statistics of the covariates. The results showed that the use of the previous covariates selection, combined with the use of an RA dataset, allowed the development of more accurate models to predict most of the attributes studied. According to Gower's general index, the RA has high similarity with the Urucu, Araracanga and Juruá blocks. However, statistics showed that increasing the distance from the RA, some relief covariates are more different. In the first chapter, the prediction models developed to predict SOCS100 showed greater accuracy and transferability than those developed to predict SOCS30. The RF algorithm generated the most accurate SOCS100 maps for the Urucu and Juruá Blocks (R2 = 0.70 and 0.51, respectively). The SOCS100 values of the maps generated for the Urucu Block region ranged from 3.89 kg C. m-2 to 10.64 kg C. m-2, while for the Juruá block they ranged from 5.03 kg C. m -2 to 10.42 kg C. m-2. In chapter 2, the best performance was also obtained with the RF algorithm in the prediction of surface and subsurface silt for the Urucu and Juruá Blocks (R2 = 0.58 and 0.52, 0.51 and 0.56 respectively). The surface and subsurface silt values of the maps generated for the Urucu Block region ranged from 208.97 g kg-1 to 576.68 g kg-1 and 215.32 g kg-1 to 517.06 g kg-1, while for the Juruá block they ranged from 236.10g kg-1 to 555.70 g kg-1 and 229.83 g kg-1 to 460.56 g kg-1, respectively. Despite the low observation density of the available dataset, the results show not only the importance of ML algorithms to map soil attributes, but also the use of specialized pedological knowledge generated in an RA to support a selection of covariates before calibrate the ML algorithms.