Research Areas
My research program is dedicated to the development of autonomous systems capable of learning from complex, high-dimensional, and multimodal data without human supervision. By bridging modern machine learning with Earth sciences, I aim to create foundation models that are scalable, generalizable, and scientifically informative.
Self-Supervised & Contrastive Learning
Problem Scope: Traditional deep learning relies heavily on large-scale labeled datasets, which are prohibitively expensive and time-consuming to acquire in domains like remote sensing or scientific time series.
Methods: I develop self-supervised learning (SSL) frameworks, including contrastive learning (SimCLR, MoCo) and masked autoencoders (MAE), tailored for non-natural image domains. This involves designing data augmentation strategies that preserve physical and spectral properties.
Scientific Contribution: These methods enable models to learn invariant and high-level representations from raw, unlabeled sensor data, significantly reducing the labeled data requirement for downstream tasks such as classification and anomaly detection.
Multimodal Representation Learning
Problem Scope: Environmental and satellite data are inherently multimodal, encompassing multispectral bands, temporal sequences, and spatial context. Conventional models often fail to capture the synergy between these diverse sources.
Methods: I research architecturally fused models that align different modalities into a unified embedding space. This includes cross-modal attention mechanisms and joint SSL objectives that leverage the complementarity of spectral and spatial information.
Scientific Contribution: My work facilitates the creation of "foundation models" for Earth observation, capable of performing zero-shot or few-shot transfer across different sensor types and geographical regions.
Spatio-Temporal Modeling & Video Understanding
Problem Scope: Modeling dynamic processes—such as vegetation phenology or urban expansion—requires capturing dependencies across both space and time, often in the presence of noise or missing data (e.g., cloud cover).
Methods: I utilize Transformers, State-Space Models (SSMs), and recurrent architectures to model long-range temporal dependencies. My research in video colorization also informs how temporal consistency can be enforced in generative tasks.
Scientific Contribution: These approaches improve the accuracy of temporal forecasting and change detection, providing more reliable data for climate monitoring and agricultural planning.
Remote Sensing & Earth Observation
Problem Scope: Interpreting satellite imagery requires addressing domain-specific challenges such as atmospheric effects, varying resolutions, and the lack of ground truth.
Methods: I apply representation learning specifically to multispectral and hyperspectral imagery, focusing on semantically-aware contrastive learning and land-use/land-cover classification.
Scientific Contribution: By developing AI tools specifically for Earth observation, I contribute to more precise environmental monitoring, enabling the tracking of biodiversity, deforestation, and agricultural productivity at a global scale.