Research
My work develops physically grounded, data-driven methods to monitor Earth's surface from space and the air. The themes below are continuing methodological threads: ideas and tools that travel across projects, sensors, and study regions. The first three describe the directions I am actively pushing today; the remaining three are the methodological foundations they rest on.
Geospatial AI for Environmental Health & Hazards
Linking remotely sensed environmental fields with spatial machine learning to study disease risk and natural-hazard regimes at scales that matter for policy.
Environmental health applications need variables that vary on ecological rather than administrative boundaries (vegetation status, surface moisture, temperature proxies, and land-cover mosaics) merged with socioeconomic indicators where available. Graph-based and spatially-aware deep models express the heterogeneous connectivity between zones and let predictors separate local forcing from regional spillovers.
A first axis of this work develops ecozone-specific predictive models for dengue outbreaks in Peru, combining vegetation indices, surface moisture, and climate with mobility and socioeconomic data to identify spatial drivers of disease hotspots. A second axis is global landslide regime analysis: using multi-sensor Earth observation and climate descriptors to characterise where events concentrate and which environmental controls recur across regions, supporting risk framing in data-scarce areas.
The emphasis throughout is on transparent covariate design and scale crossings, connecting local outbreaks and slope-scale failures to planetary envelopes of hazard.
Energy-Based Transformers in Remote Sensing
A frontier thread: replacing softmax with an energy function that supports iterative refinement, calibrated uncertainty, and physics-aware fusion.
Standard discriminative heads normalise scores through softmax and return a single forward pass. Energy-Based Transformers (EBT) instead associate each configuration (x, y) — input and candidate output — with a scalar energy E(x, y); low energy means high compatibility. This reframes inference itself.
Three mechanisms make EBTs particularly attractive for Earth observation. First, iterative refinement: small gradient or sampling moves improve a prediction while the input is held fixed, useful when SAR retrievals start from a noisy initial guess. Second, multi-modal coupling: optical and SAR features can jointly minimise a coupled energy rather than fusing only at a shared feature layer. Third, calibrated conservatism: out-of-distribution pixel types receive higher energy, instead of overconfident softmax saturation.
The promise for remote sensing is concrete. Physics-informed energy penalties keep inversions close to the manifold of plausible scattering responses (e.g. Water Cloud or radiative-transfer predictions), and test-time computation becomes a tool for global map coherence under uncertainty.
Deep Learning for Earth Observation
Building and adapting representation learners that respect scale, speckle, topology, and limited labels.
Convolutional and U-Net-style encoders remain workhorses for local textures and multi-resolution pyramids. Vision transformers and Earth-observation foundation models (Prithvi-class, SatMAE, and related self-supervised pretraining on planet-scale satellite corpora) shift the question from "train from scratch" to "transfer, fine-tune, and design task heads" that match SAR's heavy-tailed statistics or optical's cloud gaps.
Many environmental phenomena are explicitly relational: hydrology, ecological zones, and disease spillovers call for graph neural networks where nodes are regions and edges encode contiguity or covariate similarity. A cross-cutting interest is self-supervised pretraining on Sentinel time series to learn seasonally stable features before attaching sparse ground labels.
Fusion of SAR scattering with optical reflectance, the focus of theme 5 below, has consistently been a place where deep networks earn their keep, recovering biophysical parameters where either modality alone is ambiguous. In published work, a SAR–optical deep fusion approach reached ≥ 91% validation accuracy on LAI retrieval.
InSAR: Phase, Coherence & Surface Motion
Treating interferometric phase as an observable of deformation, failure, and material state, from city subsidence to mountain slopes and ice.
Interferometric SAR compares phases between satellite or airborne passes to infer line-of-sight path changes at sub-wavelength precision. In practice the science is as much about quality masks and time-series estimators as about a single interferogram: phase wrapping, atmospheric propagation, processing biases, and temporal decorrelation dominate the error budget.
My research develops end-to-end pipelines spanning baseline-aware pairing, interferogram formation and filtering, coherence-based masking, phase unwrapping, and time-series estimators (SBAS, PSI). Interpretation extends across ground subsidence, landslide creep, glacier dynamics, and co- and post-seismic displacement.
Optical–SAR Fusion for Land-Surface Biophysics
Fusing reflectance with scattering: turning two partial views of the planet into one consistent estimate.
Optical sensors emphasise canopy chemistry and photosynthetic activity (chlorophyll-linked reflectance) but suffer cloud gaps and saturate on structure at high LAI. SAR emphasises geometry, dielectric properties, and volumetric scattering, operating day or night through clouds, but its semantics are entangled with soil moisture and roughness. Fusion is therefore not just "more bands"; it is a deliberate interpolation between failure modes.
This thread spans signal-, feature-, and decision-level fusion: from physically-coupled pathways through modified Integral Equation Models and Water Cloud variants, to deep networks that learn residual corrections where linearised physics underfits. Applications include leaf area index, soil-moisture disaggregation, biomass proxies, land cover, and change detection.
Polarimetric SAR & Scattering Models
Connecting polarimetric observables to canopy and soil state through interpretable decompositions and radiative-transfer scattering models.
Full and compact polarimetry encode how electromagnetic waves interact with oriented vegetation, soil roughness, and moisture, not as a single brightness but as a partition among surface, volume, and double-bounce mechanisms. A persistent thread in this research is to stabilise mechanism separation in dual-pol and time-series settings, where classical eigen-based decompositions can become ill-conditioned, by combining optimised decomposition schemes with explicit first-order radiative transfer.
These physical models do double duty: they provide physics-aware targets and regularisers for learned predictors, and they make explicit where purely empirical mappings break, at biomass saturation, under strong soil-moisture and structure coupling, or when phenology decouples optical greenness from radar penetration.
For the full publication record, see the publications page.