Estimating and mapping tailings properties of the largest iron cluster in China for resource potential and reuse: A new perspective from interpretable CNN model and proposed spectral index based on hyperspectral satellite imagery

• The spatial distributions of TFe and SiO 2 contents of iron tailings dams were mapped. • A novel spectral index considering the absorption mechanism characteristics was proposed. • The DS algorithm reliably transferred lab-calibrated models to the GF-5 hyperspectral image. • The CNN model yielded the best predictions for the contents of TFe and SiO 2 . Iron tailings are crystalline powders predominantly composed of iron (Fe) and silicon dioxide (SiO 2 ). Spatially characterizing the physical and chemical properties of iron tailings is greatly important for optimal utilization and proper disposal of tailings. Visible-near infrared-shortwave infrared (VIS-NIR-SWIR; 350–2500 nm) spectroscopy offers a rapid, non-destructive, and cost-effective method for quantitatively analyzing tailings properties. This study aimed to quantify and map the spatial distribution of total Fe (TFe) and SiO 2 contents of tailings dams at the largest iron cluster in China using laboratory spectra and GF-5 hyperspectral images. A total of 230 samples were collected from the surface of 11 tailings dams and scanned by a VIS–NIR–SWIR reflectance spectrometer in the laboratory. A novel spectral index was developed through a multi-objective programming methodology. This novel index utilizes band ratios to identify the optimal combination of spectral bands that show a strong correlation with concentrations of TFe and SiO 2 . Simultaneously, it minimizes the impact of moisture content and particle size variations in surface tailings. In addition, the partial least squares regression (PLSR), random forest (RF) and convolutional neural network (CNN) algorithms based on laboratory spectra were used to calibrate spectral information with associated tailing properties. The contribution of wavelength in the calibration modeling process by calculating SHaply Additive exPlanations (SHAP) values. According to the results, the reflectance spectra were negatively correlated to TFe content and positively correlated to SiO 2 content. The three-band spectral index (TBI) calculated by R 827 /(R 900 × R 2200 ) correlated best to TFe with the correlation coefficient (r) of 0.87, while the R 2397 /(R 776 ×R 900 ) correlated best to SiO 2 with r of 0.70. It also minimized the effect of particle size and moisture content on the reflectance spectra of tailings properties. The CNN algorithm with laboratory spectra yielded the highest estimation accuracy for TFe (R 2 = 0.74, RPD = 1.79, RMSE = 3.69 %, LCCC = 0.74 and bias = -0.41) and SiO 2 (R 2 = 0.81, RPD = 2.15, RMSE = 1.28 %, LCCC = 0.86 and bias = −0.49). The direct standardization (DS) algorithm was applied to correct GF-5 hyperspectral image. Subsequently, the ability of TBI and the CNN model was compared for estimating and mapping the spatial distribution of TFe and SiO 2 contents based on the corrected GF-5 images. The SHAP could obtain the wavelength contribution of the CNN model in tailings spectral modeling. It can be concluded that the proposed TBI is able to rapidly characterize the spatial distribution of tailings properties, and the interpretable CNN model can provide a technical mean for accurate estimation of tailings properties based on laboratory spectra.