Identification of Granitic Pegmatites Based on GF-5 Hyperspectral Data and LSTM-Transformer Model

Lithium is a key strategic mineral to support new energy and advanced technology, mainly enriched in granitic pegmatite. The integration of hyperspectral remote sensing and deep learning frameworks has emerged as a leading approach for quantitative identification of pegmatites. However, deep learning algorithms are often limited by the scarcity of labeled samples and the challenge of extracting features from high-dimensional spectral data. In this study, a data augmentation-driven LSTM-Transformer framework is proposed for pegmatite mapping using GF-5 hyperspectral remote sensing data. In view of limited samples, this study constructed an enhanced training data set by integrating field-measured and infrared scanning spectra, simulating real spectral variability and linear mixing to improve data diversity. The proposed LSTM-Transformer architecture combines LSTM’s advantages in local sequence feature modeling and Transformer’s capacity to capture long-range global relationships, thereby enhancing the detection of weakly diagnostic spectral features. The framework was demonstrated in the Jingerquan lithium deposit, Xinjiang, China. The high-probability areas delineated by the proposed model successfully captured several pegmatite dikes documented in previous geological studies. The prediction results were further validated through field sampling. Geochemical analyses of specimens collected from the predicted high-probability target areas show an average Li 2 O grade of 7.86 wt.%, which meets industrial requirements and confirms the reliability of the proposed method. Overall, these results provide the potential of the proposed framework for pegmatite mapping and subsequent mineral exploration under the conditions of limited labeled samples.