Abstract:
Floods, cyclones, and landslides, like environmental hazards, are recurring threats in Sri Lanka.
They are increasing with climate variability. Conventional disaster forecasting systems are
facing various limitations. Use of outdated technologies, lack of high-quality data, lack of geographic
coverage, and insufficient integration of satellite images are major causes. This study
proposes an unsupervised anomaly detection framework using a CNN + LSTM Autoencoder
hybrid model to identify climate-related anomalies before disasters occur. The proposed model
integrates daily weather variables for Sri Lanka obtained from Google Earth Engine and OCHA,
together with satellite image brightness variables sourced from CIMSS. The model is trained
solely on non-disaster sequences using an 8-day sliding window to capture temporal and spatial
patterns. Therefore, the trained model is able to detect deviations from learned normalcy
through reconstruction error, allowing early warnings without labelled disaster data. The initial
evaluation yielded a moderate F1-score of 0.4,1 and the anomaly clusters aligned with known
high-risk hazard areas. When the model is retrained on a rebalanced synthetic dataset, it helps
improve accuracy to 70%. This enhanced disaster sensitivity. Finally, the retrained model was
deployed in a Flask-based web app to enable anomaly predictions from 2018 to 2025 for public
inference. This approach offers a data-efficient solution for location-based early warning
in low-resource environments while addressing the gaps in public disaster datasets. Future research
includes the use of attention mechanisms, late fusion multi-models, and integration of
high-frequency satellite images to improve accuracy and model generalisation.
