An Attention-Based LSTM Lithological Classification Using Multisensor Datasets

We propose an integrated lithological classification scheme in this study embedding a squeeze and excitation (SE) attention module in a Long Short-Term Memory (LSTM) recurrent neural network (RNN). Using high-resolution Landsat-9 and ASTER images, we integrate fieldwork and a two-stream multisource feature fusion strategy with dual-tree complex wavelet transform (DTCWT) to consolidate the intrinsic details of target features. We then leverage the synthesized dataset to train the proposed SE-LSTM model, which learns to classify target lithologies based on patterns of their feature representation within the dataset. Evaluation metrics such as Sensitivity, Specificity, Area Under the Curve (AUC), Accuracy and Confusion Matrix are used to assess the model performance. Assessing SE-LSTM’s 0.98 AUC and 84.14% accuracy against ViT, vanilla LSTM, ResNet50, KNN and RF, shows its superior efficiency and robustness in classification. SE-LSTM scores a 2.08% increase in accuracy in identifying lithology in the intricate West African Craton.