Characterization and detailed mapping of C by spectral sensor for soils of the Western Plateau of São Paulo

Abstract

Soil mineralogy and texture are directly related to soil carbon due to the physical properties of the clay surface. Traditional techniques for quantifying carbon in soil are time-consuming and expensive, making large-scale quantification for mapping unfeasible. The alternative is the use of soil sensors, such as diffuse reflectance spectroscopy (DRS), an economical, fast, and accurate technique for predicting carbon stocks. In this sense, this study aimed to (a) investigate the relationship of C with different soil mineralogical, chemical, and physical attributes for different geological and geomorphological compartments; (b) understand which spectral bands are most important for estimating C content; (c) estimate C content from diffuse reflectance spectroscopy using different mathematical techniques and indicate which one is the best for tropical soil conditions; and (d) map C contents in detail. The study area was the Western Plateau of São Paulo (WPSP), which covers approximately 13 million hectares (~ 48% of the State of São Paulo, Brazil). A total of 265 samples were collected in this area. The attributes clay, silt, sand, crystalline and non-crystalline iron, base saturation, soil density, total pore volume, total C, C stock, kaolinite/(kaolinite + gibbsite) and hematite/(hematite + goethite), hematite and goethite contents, and spectral curves were evaluated. The spectra were recorded at 0.5-nm intervals, with an integration time of 2.43 nm s−1 over the 350 to 2500-nm range (350–800 nm—visible—VIS and 801–2500 nm—near-infrared—NIR). The data were subjected to descriptive statistics, Spearman correlation, stepwise analysis, and cluster grouping for characterization purposes; partial least squares regression (PLSR) and random forest (RF) for estimation purposes; and geostatistics analysis for creation of spatial maps. Our results indicate that the highest C contents are associated with more clayey soils, oxidic mineralogy, higher total pore volume, and lower soil density in highly dissected basalt compartments. The random forest algorithm associated with the Vis–NIR spectral range is more efficient for estimating and mapping C contents. This suggests that integrating diffuse reflectance spectroscopy with machine learning techniques holds promise for shaping public policies related to land use, mitigating CO2 emissions, and facilitating the implementation of carbon credit policies in a rapid and economically efficient manner.