Geo-Data Scientist | GeoAI & Earth Observation ML | Explainable, Physics-Informed AI for Earth Systems

I build physics-aware, explainable ML that turns sparse or noisy Earth-observation, geophysical, and subsurface data into signal for climate-driven and geophysical hazards: glacial lake outburst floods, coseismic deformation, catchment flooding. A model that gets the right answer for reasons nobody can explain isn't good enough when the underlying Earth system matters as much as the prediction, so I'd rather report an ambiguous or negative result honestly than dress it up as a clean win. The six projects below are chosen to show that the approach holds up against real events, not just performs well on held-out data.

That instinct comes from where I started: at Hindalco, working with Earth-system data before I'd touched a line of ML code. Satellite and Sentinel imagery, subsurface geophysics, whatever the problem needed. An MSc in GeoData Science at Purdue (4.0 GPA) is where the ML came in, and I carried that combination into Fugro afterward, mostly in geophysical inverse theory and subsurface characterization. Five-plus years across both, now applied independently to hazard and climate problems.

Langtang Lirung PolSAR Precursor Analysis Built in the days after the catastrophic August 2026 ice-rock collapse and cascading flood on the Nepal-Tibet border (Langtang Lirung / Rasuwa-Nuwakot corridor). This started as a broader four-strand investigation: InSAR, PolSAR, Landsat, and SAR-based flood exposure mapping. The InSAR strand showed signs of phase-unwrapping ambiguity and an underdetermined control-zone network that undermined confidence in the displacement signal, so I scoped down to what actually held up: the PolSAR precursor analysis. I'd rather ship one strand I trust than four I don't.

Nepal Climate Screening: Two Independent Methods A follow-up to the PolSAR work above, asking a different question about the same event: was the pre-collapse warmth actually unusual, and does it fit the region's longer warming trend? Not proof that climate change caused the collapse, just an honestly bounded screening question, answered two different ways so the methods could check each other. A classical percentile and trend analysis on ERA5-Land placed the final week before the collapse in the 99th percentile over 76 years of history, sitting atop a real, if modest, long-term warming trend at the site. A convolutional VAE trained on regional temperature fields backed that up independently, though only after I found that reconstruction error, the usual VAE anomaly signal, was actually the wrong one for this data; KL divergence turned out to be two to three times more sensitive. The VAE also caught something the point-based method missed on its own: the sharpest regional warmth actually showed up over a week before the final spike, not during it, a two-phase pattern I only trusted after confirming it three separate ways and stress-testing the model choice across five different data splits. No permafrost data exists anywhere near the site either, which I checked directly rather than assumed, and said so plainly instead of glossing over it.

Geophysical-to-Geotechnical Inversion Toolkit Grew directly out of my time at Fugro. Rayleigh wave forward modeling (disba) feeds a global inversion (scipy differential_evolution), paired with XGBoost models that predict geotechnical properties (qt, fs) straight from geophysical data. The inversion turned up a genuine bimodal non-uniqueness result, and the tree model hit a real extrapolation artifact. I wrote both up as findings in their own right instead of quietly smoothing them over.

InSAR / Seismic Hazard Toolkit A LiCSBAS-based coseismic deformation analysis of the 2023 Kahramanmaraş earthquake sequence, built from Sentinel-1 InSAR data. The part I like: the line-of-sight velocity dipoles (±100 mm/yr either side of the rupture) and the coseismic offset time series line up almost exactly with where the fault actually moved, two independent measurements agreeing on the same answer.

Supraglacial Lake Monitor A U-Net trained on Sentinel-2 imagery to automatically detect and track supraglacial lakes on Greenland's Russell Glacier, the meltwater ponds that matter for ice sheet mass loss and, when they drain suddenly, can speed up the ice flow underneath them. Test IoU came out to 0.842 against the Qiu & Ran (2023) reference dataset, close enough that I trust it.

UK Catchment Flood & Water Security Toolkit A three-stage pipeline across two contrasting UK catchments: the Thames (low gradient, water security framing) and the Eden (steep upland, 2015 Storm Desmond). It benchmarks a conceptual linear reservoir model against an LSTM for rainfall-runoff forecasting on CAMELS-GB data, then trains leakage-free XGBoost flood-day classifiers with SHAP explainability. My favorite result: the SHAP rankings recovered each catchment's true hydrological response time straight from the data (30-day accumulated signals for the Thames, 3-day for the Eden), matching the known topography without ever being told it. A Thames-focused geospatial overlay (OSM infrastructure + Sentinel-2 NDWI via GEE) then mapped flood risk exposure concentrated around Oxford. Delivered with checkpointed models, figures, an integrated notebook, and an explicit limitations section.

Earth Intelligence Platform A modular Streamlit app that bundles eight geospatial engines (land cover, terrain, weather, and more) into one place, built on Google Earth Engine and Copernicus DEM data. The land cover classifier is the part I'm proudest of: adding SWIR bands and NDSI feature engineering, plus adaptive downsampling so it doesn't choke on large areas of interest, took the macro F1 from 0.55 to 0.60.

Carbon Verification Toolkit An independent MRV (measurement, reporting, verification) pipeline combining GEE, GEDI L4A, and OCO-2/3 data to sanity-check registry carbon credit claims, because a credit is only as good as the number behind it. Applied to the Kachung Forest Project in Uganda (CDM 4653), it landed on an independent estimate of ~324k tCO2e against the registry's claimed ~345k tCO2e, a −6.1% variance.

Urban Heat Island Prediction with Explainable ML An end-to-end GeoAI workflow predicting land surface temperature and urban heat risk for Nagpur, India, from Landsat 8/9 imagery (NDVI, NDBI, NDWI features) via Random Forest and XGBoost, with SHAP for interpretation. Built-up intensity emerged as the strongest driver of surface temperature, with vegetation showing a consistent cooling effect, a result SHAP confirmed aligns with established urban climate theory rather than merely fitting the data. Delivered as five sequential notebooks with GIS-ready GeoTIFF outputs. One of the earliest projects in this portfolio, and still one I'm fond of.

Geospatial & EO: Google Earth Engine · geemap · GeoPandas · Rasterio · rioxarray · xarray · GDAL · QGIS ML & Data Science: scikit-learn · XGBoost · LightGBM · PyTorch · SHAP · pandas · NumPy Core: Python · SQL · Git/GitHub · Jupyter · VS Code

Building this out as a portfolio of reproducible research repositories, each with its own environment, documented physical assumptions, and results reported as they came out, negative or ambiguous ones included.

Working toward doctoral research at the intersection of AI and environmental science: explainable AI for environmental hazards, physics-informed deep learning for subsurface and atmospheric systems, and Earth observation for climate resilience.

• LinkedIn: Shreya Jariwala
• GitHub: @ShreyaJari
• CV: Resume
• Email: shreyajariwala121@gmail.com