Gold Exploration using Representations from a Multispectral Autoencoder

Satellite imagery is employed for large-scale prospectivity mapping due to the high cost and typically limited availability of on-site mineral exploration data. In this work, we present a proof-of-concept framework that leverages generative representations learned from multispectral Sentinel-2 imagery to identify gold-bearing regions from space. An autoencoder foundation model, called Isometric, which is pretrained on the large-scale FalconSpace-S2 v1.0 dataset, produces information-dense spectral-spatial representations that serve as inputs to a lightweight XGBoost classifier. We compare this representation-based approach with a raw spectral input baseline using a dataset of 63 Sentinel-2 images from known gold and non-gold locations. The proposed method improves patch-level accuracy from 0.51 to 0.68 and image-level accuracy from 0.55 to 0.73, demonstrating that generative embeddings capture transferable mineralogical patterns even with limited labeled data. These results highlight the potential of foundation-model representations to make mineral exploration more efficient, scalable, and globally applicable.

💡 Research Summary

The paper presents a proof‑of‑concept framework for satellite‑based gold prospectivity mapping that leverages self‑supervised representations learned from multispectral Sentinel‑2 imagery. Recognizing that traditional field‑based exploration is costly and that labeled remote‑sensing datasets are scarce, the authors construct a massive unlabeled corpus called FalconSpace‑S2 v1.0, comprising 1,156,800 Sentinel‑2 tiles (128 × 128 px, 12 spectral bands). The dataset is deliberately diversified across acquisition dates (2020‑2025), global lat‑lon coverage, and land‑cover classes, with cloud cover limited to <30 %.

On this corpus they pre‑train a transformer‑based masked autoencoder (MAE) named “Isometric”. The encoder processes the input as 8 × 8 × 3 patch tokens, preserving both spatial context and spectral continuity. Compared with the contemporary SpectralGPT model, Isometric employs a deeper decoder and a lower masking ratio (40 % instead of the typical 90 %), which improves reconstruction quality. After pre‑training, the encoder is frozen and used as a feature extractor; its latent embeddings become the input to a lightweight XGBoost classifier.

Reconstruction performance is evaluated with five standard metrics: Mean Squared Error (MSE), Peak Signal‑to‑Noise Ratio (PSNR), Structural Similarity Index (SSIM), Spectral Angle Mapper (SAM), and Relative Global Dimensional Error (ERGAS). Isometric dramatically outperforms SpectralGPT (e.g., MSE 0.006 vs 0.062, PSNR 32.8 dB vs 21.9 dB, SSIM 0.94 vs 0.61), indicating that its latent space captures the essential spectral‑spatial information needed for downstream tasks.

For the downstream gold‑exploration task, the authors assemble 63 Sentinel‑2 images collected between June and August 2023: 33 from known gold deposits (sourced from Mindat.org) and 30 from randomly selected non‑gold locations worldwide. Each image is split into 1,024 patches of 8 × 8 × 3 pixels, yielding 33,792 gold patches and 30,720 non‑gold patches. An 80/20 split at the image level (5‑fold cross‑validation) provides training and test sets. The same XGBoost architecture is trained twice: once on raw 12‑band pixel values, and once on the Isometric embeddings.

Results show a substantial gain when using the learned representations. Patch‑level accuracy rises from 0.517 ± 0.010 (raw) to 0.681 ± 0.043 (Isometric); ROC‑AUC improves from 0.52 to 0.74, a 42 % relative increase. Image‑level accuracy climbs from 0.554 ± 0.101 to 0.733 ± 0.130, and F1‑score from 0.488 ± 0.130 to 0.729 ± 0.131. Compared with SpectralGPT, Isometric still leads by 7.2 % in patch‑level ROC‑AUC and by roughly 10 % in image‑level accuracy, demonstrating that the deeper decoder and lower masking ratio translate into more discriminative features for mineral detection. Visual inspection of RGB composites confirms that human observers cannot reliably distinguish gold‑bearing from non‑gold scenes, yet the Isometric‑based classifier correctly labels all six illustrated examples, whereas the raw‑input classifier misclassifies two of the three gold samples.

The authors acknowledge limitations: the gold dataset is small (63 images), class imbalance may affect generalization, and geographic bias could exist due to the random sampling strategy. Nonetheless, the study convincingly shows that frozen generative encoders can serve as universal feature extractors for resource exploration, even when only a handful of labeled examples are available.

Future work will expand the labeled mineral dataset, integrate additional modalities such as SAR and hyperspectral imagery, and explore multi‑temporal analysis to capture seasonal alteration signatures. The authors also suggest applying the same frozen encoder to other minerals (e.g., copper, lithium) to test transferability. In summary, the paper demonstrates that self‑supervised multispectral representation learning, combined with a simple downstream classifier, can markedly improve the efficiency, scalability, and accessibility of satellite‑based mineral prospectivity mapping.