Stacking ensemble learning algorithm based rapid inverse modelling of copper grade using imaging spectral data

Abstract

The determination of copper ore grade in a reasonably fast and accurate manner is of great practical significance for the purposes of ore dressing and ore allocation in mines. The most common method of determining the grade of copper ore is chemical analysis. However, this method has several disadvantages, including a lengthy determination period, the possibility of chemical pollution, and a lag in the results of ore dressing and ore allocation. Hyperspectral imaging technology is capable of both spectral resolution and image resolution. It is able to obtain the indicators of the sample to be measured while retaining its original physical and chemical properties. This makes it possible to overcome the shortcomings of traditional methods, allowing for accurate, non-destructive, environmentally friendly, rapid detection of samples. Stacking can often provide higher predictive accuracy than a single model by combining the predictions of multiple models, and has the advantages of reduced overfitting, model diversity, flexibility and adaptability. Stacking ensemble learning algorithm is rarely used for hyperspectral quantitative inversion modelling. In this study, 138 copper samples from the Mirador Copper Mine were employed as a data source. The spectral data of the copper samples and chemical analyses of the copper grades were collected utilising a Pika L with a Pika NIR-320 hyperspectral imager. Firstly, the raw spectral data were subjected to mutual information computation as a means of serial fusion of the spectral data, and the fused data were subjected to SG smoothing to remove noise from the spectral experiments. Subsequently, the pre-processed spectral data were subjected to feature band extraction utilising the CARS and CARS-SPA algorithms with the objective of eliminating uninformative variables and extracting valid spectral information. Finally, based on the Stacking algorithm, a highly reliable copper grade estimation model was constructed by combining various machine learning methods, and transfer learning was used to verify the accuracy and generalisation of the model. The findings of the study indicate that the feature bands selected by CARS-SPA encompass spectral ranges with sufficient chemical information, while uninformative variables are largely excluded, resulting in a notable increase in the speed and accuracy of modelling inversion operations. The Stacking ensemble learning model is more suitable for the prediction of copper grade in the Mirador copper mine compared to a single inversion model, and the CARS-SPA-Stacking inversion model has the highest accuracy, with R², RMSE, MAE, RPD, MAPE and CV reaching 0.936, 0.040, 0.019, 4.018, 0.059 and 0.267, respectively. This study is pertinent to the application of fused imaging spectral data in conjunction with the Stacking ensemble learning algorithm to copper grade inversion at the Mirador copper mine.