Quantitative inversion modeling of surface gold abundance based on remote sensing imagery and geochemical Data: An example from Tasiast-Tijirit gold district, Mauritania

Abstract

The Tasiast-Tijirit Terrane in northwestern Mauritania is an important gold mining district that mainly consists of igneous and metamorphic units that are thought to represent the remnants of older greenstone belts. Surface outcrops typically contain a high concentration of economically valuable elements. This study focuses on the quantitative inversion of auriferous soil and rock samples based on remote sensing data, highlighting the significance of using surface geochemical samples to delineate anomaly areas of gold mineralization in desert regions for effective mineral exploration programs. The backpropagation neural network inversion model was used in this work to quantitatively invert the soil and rock samples with spectral band reflectance of Landsat-7 ETM+ and GF-2 satellite imagery at 1:50000 and 1:5000 scale, respectively. Landsat-7 ETM+ was chosen because its spectral bands are almost identical to the GF-2 remote sensing data, allowing for a reasonable correlation between the datasets. Results indicate that the established model achieved R² values of modeling and test sets are 0.65 and 0.63, 0.52 and 0.49 with RMSE values of 0.009 and 0.014, 0.034 and 0.055 for soil and rock samples, respectively, using Landsat-7 ETM+. Similarly, GF-2 imagery R² values of modeling and test sets are 0.73 and 0.69, 0.60 and 0.57, with RMSE values of 0.005 and 0.004, 0.015 and 0.023 for soil and rock samples, respectively. The inversion modeling and predicted anomaly areas are well aligned with the geochemical exploration map and actual mining area. The findings suggest that although Landsat-7 imagery provides an overall distribution of surface gold elements, it is restricted in its ability to delineate high gold-rich zones in desert regions due to relatively coarse resolution besides the geological and environmental conditions such as wind erosion and weathering effects. Conversely, GF-2 imagery enabled precise delineation of the anomaly locations with rock samples, proving to be more effective owing to its higher resolution scale of 1:5000. Overall, the adopted innovative methodology that implements high-resolution satellite data with the bakpropagation neural network model promise to be very effective in enhacing minerals prediction accuracy and lowering the exploration costs.