Evaluation of regression models' accuracy for estimating soil heavy metals content from topographic rectified satellite imagery (Case study: Irankouh from Iran).

Although the traditional approach is vital in environmental studies, satellite images have recently become one of the main tools for estimating soil pollution due to their advantages. Among sources of soil pollution, heavy metals (HMs) play a crucial role; in this regard, evaluating the soil contamination anomaly is indispensable due to an enormous number of HMs mines and agricultural land irrigation by industrial sewage in developing countries. Therefore, the main goals of this study are: to evaluate Landsat-8 image capability for estimating HMs content (Cr, Cd, Co, and AS) of the soil, performance evaluation of linear and nonlinear regression models, and topographic influence on the estimation accuracy of heavy metals. To do so, we applied multivariate-linear, partial least square, exponential regression, and neural network models to achieve a relation between the HMs content and soil spectral reflectance. Regarding the R2 value of models, the neural network demonstrated a more acceptable performance than other models. The results show that topographic correction improves the accuracy of HMs estimation; for example, the Minnaert model obtained the highest score (equal to 30.02), which is sufficient to support the preceding sentence. Furthermore, the neural network and Minnaert (with a score of 17.56) would produce the best results for estimating HMs content. [ABSTRACT FROM AUTHOR]