GeoAI / Remote Sensing Direction
Domain-Generalized GeoAI
Domain-generalized multimodal GeoAI for climate-smart agriculture.
This emerging research direction extends my reliability-focused computer vision and multimodal learning work toward remote sensing and agricultural monitoring under geographic, temporal, climate, and sensor shift.
Overview
The goal is to study satellite-based AI models that remain useful across unseen regions, years, climate regimes, crop systems, and sensor conditions. This page presents an active direction rather than a completed peer-reviewed project.
Problem Statement
Many agricultural remote-sensing models perform well under random train-test splits but fail in realistic deployment settings where geography, crop calendars, weather, soil background, clouds, and sensor conditions change.
Motivation
Climate-smart agricultural monitoring needs models that generalize beyond the data distribution on which they were trained. This matters for crop mapping, early stress detection, drought and flood impact monitoring, irrigation assessment, yield-risk estimation, and food-security decision support.
Methodology Plan
I am exploring selective domain residualization and reliability-aware SAR-optical-weather fusion. The aim is to preserve transferable crop and stress evidence while reducing dependence on unstable domain-specific shortcuts.
- Separate stable crop and land structure from nuisance domain information.
- Fuse Sentinel-1 SAR, Sentinel-2 optical imagery, weather data, and environmental covariates.
- Evaluate generalization under region, year, climate-zone, and sensor-condition shifts.
Target Data Modalities
| Sentinel-1 SAR | Structure and moisture-related information, including cloudy-condition utility. |
|---|---|
| Sentinel-2 optical | Vegetation spectral response and phenological information. |
| Context variables | Weather, terrain, and environmental covariates for climate and land-context reasoning. |
Evaluation Plan
- Leave-region-out generalization.
- Leave-year-out generalization.
- Leave-climate-zone-out generalization.
- Missing-modality and degraded-sensor stress tests.
- Calibration, uncertainty, and worst-domain behavior analysis.
Current Status
This is an emerging research direction. Results, datasets, figures, code, and citations will be added only after they are ready to share.
My Role
My role is to formulate the domain-generalization problem, design the multimodal reliability framework, implement experiments, and analyze robustness across realistic deployment-style splits.