Free iron oxide content in tropical soils predicted by integrative digital mapping

The free iron content is a vital indicator of pedogenic processes in tropical soils and can be used to understand the soil's weathering history and aid in classification. Despite its importance in agriculture and pedology, laboratory analyses of soil iron content are not widely used because they are costly and time-consuming. Remote sensing data combined with digital soil mapping are effective tools to regionalise soil iron content. They can reduce the number of soil samples needed to characterise soil variability and, consequently, laboratory analysis costs. This study aimed to create a strategy for mapping free iron content using a 35-year time series of Landsat images combined with topographic parameters at two spatial resolutions (5 and 30 m) in a region with high variability in soils and geology in the state of Sa & SIM;o Paulo, Brazil. The dataset comprised 344 observations of free iron content at 0-20 cm depth over a 2574 km(2) area. The dataset was split into calibration and a validation set (85:15%), and the environmental variables were chosen based on the scorpan factors. Spatial prediction functions for free iron were developed using several machine learning algorithms linking soil observations with the environmental variables. We found that the temporal bare soil image improved model performance. Although 5 and 30 m resolution terrain data differed slightly, the best-fit model was obtained at 5 m resolution (root mean square error, 25.09 g kg(-1); adjusted R-2, 0.84). Among the evaluated machine learning algorithms, Random Forest was the most accurate method for predicting free iron distribution in the study area. The free iron content map can identify soil types in more detail and should be prioritised in future pedological studies.

Automated summary

Free iron content is an important indicator in tropical soils, and this study successfully developed a mapping strategy for predicting free iron distribution using remote sensing data, digital soil mapping, and machine learning algorithms.