Enhancing land use and land cover mapping with uncertainty-aware 3D-UNet and Squeeze and Excite attention mechanism

Deep learning models have significantly improved Land Use and Land Cover (LULC) mapping, but integrating and quantifying uncertainties remain a challenge. This paper presents a novel 3D-UNet approach that incorporates uncertainty quantification (UQ) and attention mechanisms to enhance LULC mapping accuracy by providing valuable insights into model confidence. The integration of Squeeze and Excite (SE) attention enables the model to focus on the extraction of the most informative features, ensuring the effective representation of critical land features. Monte Carlo Dropout is employed to generate both aleatoric and epistemic uncertainty maps, which capture the inherent uncertainty in the data and the model predictions. These uncertainty maps are integrated with satellite images to provide a more reliable and accurate LULC classification. The results show that the proposed model outperforms traditional methods in LULC segmentation, achieving an average improvement of 4% across key evaluation metrics, including accuracy, precision, recall, F1-score, Hausdorff distance, and Dice score, when compared to the baseline. The improvements in Hausdorff distance and Dice score indicate better delineation of land boundaries and a more accurate representation of land cover classes. These results highlight the robustness and reliability of the proposed approach, underscoring its potential in applications such as precision agriculture.