HiRISE Band Reconstruction and Data Processing

HiPredict is a deep learning and scientific data processing framework for reconstructing missing image data in high-resolution Martian imagery from the HiRISE instrument on Mars Reconnaissance Orbiter, motivated by the intermittent failure of the central RED4 detector since 2023.

• End-to-end workflows from raw IMG files through calibration, stitching, co-registration, and tiling.
• HiPredict was trained on over 40,000 images to reconstruct missing central-detector data in HiRISE observations.
• The model reduced per-pixel normalised mean absolute error by 71.21% relative to the SYN4 baseline and reduced mean bias from -0.205% to -0.005%.

Conference abstract: Lunar and Planetary Science Conference 2026

Publication in progress

Reinforcement Learning for Atmospheric Balloon Station Keeping

Imperial College Space Society project exploring reinforcement learning for atmospheric balloon station keeping, modelled on Google's stratospheric balloon initiative. I serve as AI Lead for the 2025-26 team.

• Developed a custom reinforcement learning environment for altitude-based control and long-duration station keeping.
• Built around atmospheric dynamics and weather-driven motion relevant to stratospheric balloon systems.
• Part of Project Stratus, an Imperial College Space Society effort to explore autonomous balloon navigation.

Project repository: Project Stratus GitHub

Supervised Water-Ice Detection on the Moon

MSc thesis project on predicting water-ice presence at the lunar poles using supervised learning and multimodal remote sensing data. The work combined Diviner, LOLA, M3, and Mini-RF products to score points across the polar surface, extending beyond permanently shadowed regions alone.

• Built a full data pipeline to decode, process, interpolate, and combine heterogeneous lunar datasets into ML-ready features.
• Developed a fully connected neural network for polar-scale water-ice scoring from multimodal orbital data.
• The resulting resource-potential maps showed strong precision (0.7979), while also identifying that only 57.8% of permanently shadowed terrain exhibited detectable ice signatures.

Conference poster: Space Resources Week 2025