A Novel Scheme for Mapping of MVT-Type Pb–Zn Prospectivity: LightGBM, a Highly Efficient Gradient Boosting Decision Tree Machine Learning Algorithm

• Published in: Natural resources research (New York, N.Y.) N.Y.), 2023-12, Vol.32 (6), p.2417-2438
• Main Authors: Hajihosseinlou, Mahsa, Maghsoudi, Abbas, Ghezelbash, Reza
• Format: Article
• Language: English
• Subjects: Algorithms; Chemistry and Earth Sciences; Computer Science; decision support systems; Decision trees; Earth and Environmental Science; Earth Sciences; Fossil Fuels (incl. Carbon Capture); Geography; Iran; Lead; Learning algorithms; Machine learning; Mapping; Mathematical Modeling and Industrial Mathematics; memory; Mineral Resources; Original Paper; Performance indices; Physics; prediction; Remote control; Remote sensing; Sampling methods; Statistics for Engineering; Sustainable Development; Tectonics; Zinc
• ISSN: 1520-7439; 1573-8981
• DOI: 10.1007/s11053-023-10249-6

The gradient boosting decision tree is a well-known machine learning algorithm. Despite numerous advancements in its application, its efficiency still needs to be improved for large feature dimensions and data capacities. Each feature necessitates a time-consuming scan of all samples to determine the estimated information gain of all split points. To address this problem, a light gradient boosting machine (LightGBM) algorithm was introduced here. It was applied to construct a Pb–Zn prospectivity model for the Varcheh district, west of Iran. The technique utilizes the histogram optimization approach, allowing quicker training and more effective memory use, the leaf-wise strategy to prevent overfitting, the exclusive feature bundling technique to minimize the features' dimension, and the gradient-based one-sided sampling method to lower the data dimension. The inputs for the model included geological layer, tectonic control, geochemical data, and remote sensing evidence. It is worth noting that parameter tuning for the LightGBM technique is challenging. The important hyperparameters were therefore improved through GridSearchCV. The proposed model was compared to extreme gradient boosting (XGBoost) algorithm as an ensemble machine learning algorithm. A confusion matrix was utilized to compare the overall performance of the two prediction algorithms. The LightGBM had a lower error as well as greater accuracy, precision, and F1_measure and thus outperformed XGBoost based on all performance indices. Most importantly, LightGBM had predicted just 10% of the research area to be prospective, delineating 92% of the know Pb–Zn deposits. Overall, the LightGBM has provided novel insights into the effective and rapid processing of large geodatasets for mineral prospectivity mapping.