Martin Wright

Coastal Camera Image Processing

Development of Python-based image processing tools for a nine-camera coastal monitoring network. The system applies the pinhole camera model with Caltech lens distortion, robust homography estimation (MAGSAC++) for frame stabilisation, and backward-warped oblique-to-plan rectification using surveyed ground control points. Temporal pixel instruments extracted from rectified imagery enable quantitative measurement of beach morphology change.

Stochastic Co-variant Rainfall Extension

A novel approach to extending short climate records using extreme value theory and conditional probability modeling. This method extends 52-year Virtual Climate Station rainfall records to 170+ years by leveraging historical gauge data, operating at the 19-month timescale critical for drought characterization. The approach combines Generalised Pareto Distribution tail modeling with empirical resampling and analog pattern matching to preserve both extreme event characteristics and seasonal variability.

Intelligent Semantic Caching for Video Surveillance Systems

This paper introduces a novel hybrid framework for semantic video retrieval and localized caching in surveillance systems. This approach employs a two-tiered computer vision strategy using a real-time YOLO model deployed at the network edge combined with OpenAI's CLIP, where embeddings are generated to capture rich, high-dimensional zero-shot semantic content. These embeddings are then synchronized with a cloud-based vector database to establish a semantically indexed reservoir for query processing. Finally, we leverage a Large Language Model (LLM) augmented with Retrieval-Augmented Generation (RAG) to process natural language queries and index cached frames.

LSTM Drone Scheduling Model

This report details the development of a dual Long Short-Term Memory (LSTM) neural network model to optimize drone-based coastal surveys for marine species monitoring, which are frequently hampered by adverse weather. Trained on historical time-series data, the model forecasts conditions at two sites and uses a weighted policy to recommend the better location for drone flights, minimizing wasted efforts and improving the effectiveness of ecological data collection.

VAE for Hydrological Stochastic Generation

Contemporary methods for stochastically generating climate-change impacted rainfall patterns involve site-based stochastic methods. This paper builds on recent IBM Research methods using VAE models to better model co-variant rainfall patterns by using a prior-posterior sampling method based on latent space.