Mo-BAPER: A Modified TabNet Employing Global Average Pooling for Bogor Area Landslide Prediction

In this study, Mo-BAPER, a modified TabNet classifier, was developed to improve the accuracy of landslide prediction in the Bogor region of Indonesia. The dataset used was provided by the Bogor City Regional Disaster Management Agency (BPBD Kota Bogor) and consisted of 1,644 balanced data points, with 822 landslide events and 822 non-landslide cases. Initially, a standard TabNet model was utilized and optimized through hyperparameter tuning, resulting in a cross-validated accuracy of 82.29%. Additional metrics included a mean sensitivity of 83.09%, mean specificity of 81.48%, mean precision of 82.47%, mean recall of 82.29%, and a mean F1 score of 82.26%. A modification was introduced to improve the model's performance further. The TabNet model was used as a feature extractor, and its embeddings were passed through a global average pooling (GAP) layer followed by a neural network with fully connected layers. This modification significantly improved the model's generalization ability, resulting in an accuracy of 92.62%, sensitivity of 92.65%, specificity of 92.59%, precision of 92.77%, recall of 92.62%, and F1 score of 92.61%, demonstrating its effectiveness in landslide prediction.