<p>Natural hydrogen (H<sub>2</sub>) emerges as a clean energy resource generated by different geological processes in multiple geological contexts. Current hydrogen exploration activities rely mainly on surface geochemical surveys focused on diagnostic morphological expressions. Analogous to the early stages of hydrocarbon exploration, these first-pass indicators are used to delineate prospective corridors and rank targets, thereby de-risking and guiding subsequent subsurface investigations. However, these surface expressions remain elusive due to sparse sampling and the absence of robust detection workflows. This study presents a machine learning-based remote sensing technique for pixel-wise hydrogen prospectivity mapping applied on the Atyrau region in the southeastern part of the petroliferous Pricaspian Basin, western Kazakhstan. We utilize Sentinel-2 Level-2 A imagery and extract 13 spectral bands (B1–B12, B8A), 5 spectral indices (e.g., NDVI, NDWI), and 4 texture features, forming a 22-dimensional input feature space. Four classifiers, Random Forest (RF, <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(\text{n}=200\)</EquationSource> </InlineEquation>), eXtreme Gradient Boosting (XGB, <InlineEquation ID="IEq2"> <EquationSource Format="TEX">\(\text{n}=100\)</EquationSource> </InlineEquation>, max depth = 5), Stochastic Gradient Descent (SGD, loss = log), and LightGBM (LGBM, learning rate = 0.1), were trained on 120,560,400 labeled pixels with 4,621,344 positive labels derived from high-confidence expert polygons. Performance metrics revealed strong recall for XGB (91.8%) and LGBM (93.1%), albeit with lower precision (6.8% and 6.4% respectively), reflecting class imbalance and rare-event detection challenges. The ROC AUC ranged from 0.50 (SGD) to 0.81 (XGB), with Precision-Recall AUC peaking at 0.17 for RF. Explainable AI techniques based on SHAP (SHapley Additive Explanations) were applied to quantify pixel-wise feature contributions for each classifier. SHAP attributions, overlaid on geologic cross-sections, revealed spatial correlations with known faults and surface anomalies. Final predictions, rasterized into labeled GeoTIFFs at 10 m resolution, indicate locations of anomalies. These promising screening-type exploration results allow for narrowing down areas for direct surface measurements and focusing on understanding the hydrogen system and play concepts. This demonstrates that workflow provides a scalable, interpretable, and cost-effective methodology for natural hydrogen exploration and can be adapted to other data-sparse basins worldwide.</p>

Integrated pixel-wise remote sensing and explainable machine learning for natural hydrogen exploration in southeastern part of Pricaspian Basin, Western Kazakhstan

期刊论文
出版模式：

开放获取
发表日期：

2026年2月26日

查看全文

Dennis Delali Kwesi Wayo,
Leonardo Goliatt,
Randy Hazlett,
Milovan Fustic,
Mahmoud Leila

摘要

Natural hydrogen (H₂) emerges as a clean energy resource generated by different geological processes in multiple geological contexts. Current hydrogen exploration activities rely mainly on surface geochemical surveys focused on diagnostic morphological expressions. Analogous to the early stages of hydrocarbon exploration, these first-pass indicators are used to delineate prospective corridors and rank targets, thereby de-risking and guiding subsequent subsurface investigations. However, these surface expressions remain elusive due to sparse sampling and the absence of robust detection workflows. This study presents a machine learning-based remote sensing technique for pixel-wise hydrogen prospectivity mapping applied on the Atyrau region in the southeastern part of the petroliferous Pricaspian Basin, western Kazakhstan. We utilize Sentinel-2 Level-2 A imagery and extract 13 spectral bands (B1–B12, B8A), 5 spectral indices (e.g., NDVI, NDWI), and 4 texture features, forming a 22-dimensional input feature space. Four classifiers, Random Forest (RF, \(\text{n}=200\) ), eXtreme Gradient Boosting (XGB, \(\text{n}=100\) , max depth = 5), Stochastic Gradient Descent (SGD, loss = log), and LightGBM (LGBM, learning rate = 0.1), were trained on 120,560,400 labeled pixels with 4,621,344 positive labels derived from high-confidence expert polygons. Performance metrics revealed strong recall for XGB (91.8%) and LGBM (93.1%), albeit with lower precision (6.8% and 6.4% respectively), reflecting class imbalance and rare-event detection challenges. The ROC AUC ranged from 0.50 (SGD) to 0.81 (XGB), with Precision-Recall AUC peaking at 0.17 for RF. Explainable AI techniques based on SHAP (SHapley Additive Explanations) were applied to quantify pixel-wise feature contributions for each classifier. SHAP attributions, overlaid on geologic cross-sections, revealed spatial correlations with known faults and surface anomalies. Final predictions, rasterized into labeled GeoTIFFs at 10 m resolution, indicate locations of anomalies. These promising screening-type exploration results allow for narrowing down areas for direct surface measurements and focusing on understanding the hydrogen system and play concepts. This demonstrates that workflow provides a scalable, interpretable, and cost-effective methodology for natural hydrogen exploration and can be adapted to other data-sparse basins worldwide.