Spatial prediction of soil surface texture in a semiarid region using random forest and multiple linear regressions

Abstract

Soil texture is an essential and extremely variable physical property that strongly influences many other soil properties that are highly relevant for agricultural production, e.g., fertility and water retention capacity. In plain areas, terrain properties derived from a digital elevation model are not effective for digital soil mapping, and the variation in the properties of such areas remains a challenge. In this regard, remote sensing can facilitate the mapping of soil properties. The purpose of this study was to evaluate the efficiency of using of data obtained from the Thematic Mapper (TM) sensor of Landsat 5 for digital soil mapping in a semi-arid region, based on multiple linear regression (MLR) and a random forest model (RFM). To this end, 399 samples of the soil surface layer (0–20cm) were used to predict the sand, silt and clay contents, using the bands 1, 2, 3, 4, 5 and 7, the Normalized Difference Vegetation Index (NDVI), the grain size index (GSI), and the relationships between bands 3 and 2, bands 3 and 7, and bands 5 and 7 (clay index) of the Landsat 5 TM sensor as covariates. Among these covariates, only band 1 (b1), the relationship between bands 5 and 7 (b5/b7) for sand, silt and clay, and band 4 (b4) for silt were not significantly correlated according to Pearson's correlation analysis. The validation of the models showed that the properties were best estimated using the RFM, which explained 63% and 56% of the spatial variability of sand and clay, respectively, whereas the MLR explained 30% of the spatial variation of silt. An analysis of the relevance of the variables predicted by the RFM showed that the covariates b3/b7, b5, NDVI and b2 explained most of the variability of the considered properties. The RFM proved to be more advantageous than the MLR with respect to insensitivity to overfitting and the presence of noise in the data. In addition, the RFM produced more realistic distribution maps of the soil properties than did the MLR, taking into account that the estimated values of the soil attributes were in the same range as the calibration data, while the MLR model estimates were out of the range of the calibration data.