Uncertainty-driven lithological mapping: from data uncertainty to model uncertainty

Traditional lithological mapping with remote sensing data and deep learning models often ignores the sources of uncertainty arising from:Data uncertainty — due to noise, limited observation, sampling issues, and measurement errors.Model uncertainty — due to stochasticity in model parameters, architectures, or optimization. This code investigates uncertainty quantification in lithological mapping using remote sensing data and a Bayesian U-Net model:(1) Data uncertainty quantification via stochastic simulation of probabilistic labels and noisy input images.(2) Model uncertainty quantification via a Bayesian U-Net fully convolutional neural network (U-Net FCN) incorporating Monte Carlo Dropout.(3) Visualizing uncertainties using SHAP technology to enhance model explainability. InstallationPython ≥ 3.7TensorFlow ≥ 2.1 Key Scripts1. unet_data_uncertainty_quantification.pyQuantifying data uncertainty through simulating noisy inputs and probabilistic labels.Main features:Generates multiple soft label realizations (generate_soft_labels()).Trains U-Net models across NUM_ITERATIONS simulations.Injects Gaussian noise to evaluate prediction stability.Outputs:soft_labels_collection.mat — generated soft_labels (*.mat).uncertainty_predictions_cube.mat — a cube of predictions for uncertainty estimation (*.mat).training_log.csv — per-iteration training metrics. 2. unet_model_uncertainty_quantification.pyQuantifies model uncertainty using Monte Carlo dropout.Main features:Builds a dropout-regularized U-Net (build_model()).Performs Monte Carlo sampling during inference (evaluate_model_mc_dropout()).Outputs:mc_pred_map.mat — multiple Monte Carlo-sampled predictions.Training logs (training_metrics.csv)— per-iteration training metricsVisualization of mean prediction and uncertainty map. Case study data14-bands ASTER data in VNIR-SWIR-TIR regions. image_train.tif — training image.class_train.tif — training label.image_test.tif — testing image. Author & ContactAuthor: Ziye WangAffiliation: China University of GeosciencesContact: ziyewang@cug.edu.cn