Internship Projects
2025-2026
Adapting Pre-trained Deep Learning Models with Limited Data for Post-Disaster Damage Classification: A Case Study of Cyclone Gabrielle, New Zealand
This project adapted pre-trained deep learning models to identify disaster-related damage from aerial imagery using only a small amount of labelled data. Using pre- and post-disaster imagery from Cyclone Gabrielle, several pre-trained models were compared and fine-tuned to classify areas as damaged or undamaged, with the best model reaching 97.67% accuracy. Knowledge distillation then produced a smaller model that retained high damage-detection performance while cutting parameters by roughly 91%, supporting faster disaster assessment and more efficient allocation of resources in affected areas.
Chanuka Danansooriya
AI-Driven Flood Intelligence System for Historical Flood Impact Data in New Zealand
This project develops an AI-powered Flood Intelligence System that uses Large Language Models to extract and structure historical flood impact information from New Zealand news reports — locations, severity, human impacts, infrastructure damage and economic losses. The system turns unstructured reporting into an AI-ready dataset with an interactive dashboard, supporting natural hazard research, disaster risk assessment and evidence-based decision-making to strengthen flood resilience in New Zealand.
Omkar Joshi
Exploring Computer Vision Techniques for Detecting Coastal Cliff Instability
This project investigated machine learning and computer vision methods for semi-automated coastal cliff change detection using citizen science imagery. As part of a wider PhD research programme at Massey University and through research internships at CRISiSLab, the team explored feature-matching and image-registration methods — including SIFT, ORB, LoFTR and DINOv2 — alongside automatic Ground Control Point (GCP) transfer and camera geometry processing. They also investigated prompted semantic segmentation using SAM and LangSAM, and image-based change detection combining SSIM with deep-learning visual representations. The work contributed experimental workflows for extracting information about coastal cliff instability from repeat photographs, with findings presented in two international conference papers.
Hussain Raza
Meenambika
Chandirakumar
Supun Gamlath
Mobile Application Development for an AI-Powered On-Site Earthquake Early Warning System
This project covers the design and development of a mobile application for an AI-powered, low-cost, on-site Earthquake Early Warning (EEW) system built on Raspberry Shake seismic sensors. The application automatically configures and connects to the sensor, visualises seismic waveforms in real time, and monitors sensor condition including CPU usage, storage, station information and location. When the AI-based detection system identifies a potential earthquake, the application delivers immediate alerts through low-latency UDP communication while active, and through Firebase Cloud Messaging (FCM) when minimised or closed. It also provides general earthquake safety and preparedness information to support users during seismic events.
Janith Aththanayaka
Weatherly: From Weather Forecasts to Personalised Action through Generative AI-Powered Impact-Based Early Warning
Weatherly investigates how weather forecasts can be translated into personalised, activity-specific guidance to support informed planning and early action. The project combines forecast information with the timing, location and type of a user's planned activities, and explores generative AI for interpreting activity descriptions and proposing alternative arrangements. Forecast-based checks and explicit user confirmation constrain any proposed changes, keeping users in control of their decisions. The intended outcome is a research prototype and an evaluation framework for assessing whether contextual, AI-assisted guidance improves decision quality compared with forecast-only and non-AI approaches.
Kaviru Herath
Raspberry Shake based EEW using SeisComP as a centralized processing platform
This project aims to develop a centralized Earthquake Early Warning (EEW) platform using distributed Raspberry Shake seismic sensors and SeisComP for real-time seismic processing. Live waveform data from multiple stations are transmitted to a central server, where SeisComP performs automatic phase picking, earthquake detection, localization, magnitude estimation, and event generation. The system is designed to provide rapid and reliable earthquake information while evaluating detection accuracy and processing latency under different sensor-network configurations. The expected outcome is a scalable backend platform that can support real-time earthquake monitoring and early-warning applications.
Nilesh Ravishan
“Terrasense: An End-to-End Real-Time Earthquake Early Warning and Seismic Monitoring Platform”
This project presents Terrasense, an end-to-end seismic monitoring and earthquake early warning platform combining a reactive microservice backbone with specialized client and administrative web interfaces. Its central objective is to capture continuous, high-rate UDP telemetry from distributed edge seismometers, persist massive time-series waveforms, and broadcast live alerts and visualizations with sub-second latency. While the event-driven NestJS backend orchestrates data processing via Apache Kafka, TimescaleDB, and WebSockets, the client frontend empowers public and organizational users to monitor live stations, compare waveforms, and export historical data. Coupled with an administrative dashboard for role-based governance, sensor metadata management, and real-time system observability, the platform provides a complete, scalable solution for community-level earthquake preparedness.
Bimsara Janakantha
Rapid landslide detection using open-source geospatial data: A Google Earth Engine-based machine learning approach in Sri Lanka.
This study developed an efficient, scalable landslide detection methodology that integrates open-source geospatial data and machine learning techniques in Google Earth Engine, focusing on model generalisation and transfer-learning concepts in Sri Lanka. The gradient tree boost (GTB) technique performed best, demonstrating strong model generalisation. The model (C-GTB-M4), which combines optical, topographic, and synthetic aperture radar data, showed superior spatiotemporal transferability in transfer learning. The results show that the developed workflow reliably and efficiently detects rainfall-induced landslides in Sri Lanka and can be applied to similar data-scarce regions.
Sameera Lasantha