Session summary
This GeoField 2026 methods session examines how Earth observation data become usable evidence for geospatial impact evaluation. The presentations focus on the measurement, processing, modeling, and interpretation choices that shape geospatial indicators before they are used in causal analysis.
Moderator: Kunwar K. Singh
The Uses (and Misuses) of Earth Observation Data on Weather and Vegetation
Jeff Michler and Elinor Benami present guidance on using Earth observation data to measure weather, vegetation, and environmental shocks. The presentation highlights how different rainfall, temperature, and vegetation products can produce different values for the same place and time because they rely on different sensors, models, assumptions, and correction procedures. It also cautions that common indicators such as NDVI, drought measures, and flood indicators should be matched carefully to the mechanisms being studied, validated where possible, and tested across alternative products and definitions.
Geospatial Analytics and Advanced Methods for Impact Evaluation
Sara Sayedi presents a chapter on geospatial analytics and advanced methods for impact evaluation. The presentation covers key workflow choices including preprocessing, analysis-ready data, cloud masking, scale, aggregation, sensor fusion, interpolation, feature extraction, machine learning, deep learning, and cloud computing. The chapter emphasizes that these technical choices shape the variables used in evaluation and should be documented, versioned, validated, and interpreted with attention to uncertainty.
Geospatial Foundation Models for Data Generation and Indicator Production
Vivek Sakhrani presents an introduction to geospatial foundation models and their potential role in data generation and indicator production. The presentation discusses when foundation models may be useful, especially in data-scarce settings where labels are limited, existing products are inadequate, or researchers need to generate features across regions. It also introduces embeddings, change detection, classification, segmentation, and temporal reconstruction, while cautioning that foundation model outputs require validation, uncertainty assessment, and transparency before being used in impact evaluation.
Together, the session provides a bridge between geospatial data production and causal analysis. Earth observation can greatly expand what researchers can measure, but credible geospatial impact evaluation depends on understanding how indicators are constructed, where uncertainty enters, and how measurement choices affect causal claims.