Optimizing U-Net for Landslide Detection: Impact of Batch Size and Overfitting

Landslides pose significant risks to infrastructure and human life, making their early detection crucial for disaster prevention. Traditional landslide mapping methods rely heavily on manual interpretation and heuristic approaches, which are time-consuming and prone to inaccuracies. To address this, we implement a deep learning-based U-Net model for automated landslide detection using satellite imagery from the Landslide4Sense dataset. The dataset consists of multi-spectral images (128×128×14) and ground-truth masks, enabling pixel-wise classification. Our approach includes rigorous data pre-processing, normalization, and a batch size optimization study, exploring the effect of varying batch sizes on model performance. The U-Net architecture is trained using binary cross-entropy loss and the Adam optimizer, with custom callbacks implemented for automated model tracking and performance logging. The model achieves a 98.56% accuracy, but an analysis of precision (63.3%) and recall (56.0%) reveals overfitting issues, limiting its generalization to unseen data. To address this, we conduct an extensive overfitting analysis, evaluating metric trends across epochs, including accuracy, precision, recall, F1-score, and the precision-recall curve. Our results indicate that batch size selection significantly impacts model performance, influencing both recall sensitivity and false positive rates. While our model outperforms many existing methods in accuracy, it lags in precision-recall balance, suggesting a need for improved generalization strategies. This study highlights the importance of hyperparameter tuning and in-depth performance tracking in landslide detection models. We experimented with dropout regularization and batch normalization but observed limited improvement in recall. he model was optimized using batch size tuning and learning rate adjustments to mitigate overfitting. However, the model still exhibited precision-recall imbalance, indicating scope for improved dataset diversity and hyperparameter tuning. Our findings provide valuable insights into optimizing deep learning models for real-world disaster detection applications.