Deep learning classification of elevated gamma-radiation zones using Sentinel-2 spectral data

Abstract

In a previous study, the Normalized Difference Gamma-Ray Index (NDGRI) was introduced as a spectral approach to highlight surface conditions associated with elevated gamma radiation. When validated against car-borne and airborne spectrometry surveys, NDGRI demonstrated potential for delineating radiation anomalies in semi-arid terrains. Detecting such anomalies from satellite data is of practical importance, as it offers a cost-effective way to narrow down prospective zones before deploying expensive car-borne or airborne surveys. Building on this foundation, this study assesses whether deep learning can extend this approach beyond index-based methods by classifying elevated radiation from Sentinel-2 spectral data. Multilayer perceptron (MLP) networks were implemented for per-pixel spectral classification, using fully connected dense layers with dropout regularization. Hyperparameters, including number of layers, neurons per layer, dropout rates, optimizers, and learning rates, were tuned via Keras Tuner and AutoKeras in Python to balance model complexity and generalization. Training and validation samples were extracted from car-borne and air-borne gamma-ray spectrometry heatmaps, providing pixel-level labels for supervised learning. Three input configurations were tested: NDGRI provided as a single-band raster, the full set of 11 Sentinel-2 bands, and selected band subsets that showed diagnostic potential in our previous NDGRI analysis. The results showed that the NDGRI raster performed strongly as a compact predictor of elevated radiation, whereas multi-band inputs provided better overall performance, with certain subsets consistently outperforming NDGRI alone. These findings indicate that although NDGRI efficiently captures key spectral contrasts, the inclusion of targeted spectral combinations enables the network to exploit complementary features and improve classification robustness. The findings suggest that deep learning can serve as a powerful extension of NDGRI-based workflows, combining the interpretability of targeted indices with the flexibility of data-driven classification. This approach could offer a scalable and computationally efficient means of mapping elevated gamma-radiation anomalies from multispectral satellite imagery, with implications for cost-effective uranium exploration in semi-arid terrains.