Visible-spectrum image analysis for copper concentration measurement in heap leach solutions: Characterization of photographic and model parameters using a laboratory prototype

Abstract

Heap leaching is widely used for the recovery of copper from low-grade oxide ores. Continuous monitoring of copper concentration in the pregnant leach solution (PLS) is essential for tracking leaching kinetics, determining optimal irrigation termination, and balancing metallurgical inventories at the individual-heap level; however, measurement at heap drainage channels remains operationally challenging owing to limited accessibility and reliance on manual sampling with delayed laboratory analysis. This study characterizes the photographic and modeling parameters of a laboratory prototype that estimates copper concentration in copper sulfate solutions using empirical, image-based colorimetric regression with low-cost consumer imaging hardware rather than a laboratory spectrometer, as a basis for future in-line sensing. Synthetic PLS solutions (0 to 40 g Cu/L in sulfuric acid) were imaged under controlled illumination (fixed-color, fixed-intensity LED light, diffused inside a dark chamber and not collimated) using a Nikon D3100 digital single-lens reflex (DSLR) camera across four channels (red, green, blue, and a white composite) at three intensity levels. For each channel, an absorbance-like feature was derived from the channel intensity relative to a zero-copper blank. A dataset of 432 images was subdivided into 256 spatial sub-samples per image across four channels, generating 442,368 channel-level observations, which were reduced to 15,120 physically coherent observations by sequential filtering (Kendall rank correlation, then linear and quadratic absorbance-concentration bounds). A feedforward artificial neural network (ANN) predicted copper concentration from the four color features and was benchmarked against a multivariate linear model used as a Lambert-Beer baseline. On a separate simulated, controlled dataset, the ANN achieved an RMSE of approximately 0.55 g/L (residual standard deviation of 0.40–0.41 g/L), a reduction of approximately 37 % in residual standard deviation relative to the best linear baseline (from 0.64 to 0.40 g/L) on simulated data. Experimental validation yielded a prediction error of approximately 25 %. This error is structured rather than random, arising from intensity-level displacement caused by specular reflections on the cylindrical sample cell, and is comparable to the ∼ 20 % uncertainty of current manual sampling; it is therefore a hardware and data-acquisition limitation rather than a limitation of the sensing concept. Under real PLS conditions, the error would be expected to increase, so the laboratory value should be read as a best-case bound. A flat-window flow cell, collimated illumination, and inclusion of the recorded illumination intensity as a model input are identified as the priority improvements.