Developing Reliable Machine Learning Methods for Earth Observation and Landscape Change

Abstract

Machine learning (ML) and artificial intelligence (AI) models have become popular research tools for geospatial analysis, but the standard AI/ML metrics that guide model development do not necessarily show whether an output supports the scientific task. The difference between evaluation benchmarks and downstream utility creates a ``reliability gap'' between developers and users of geospatial AI workflows. This dissertation develops methods that incorporate ``reliability'' into geospatial AI workflows in multiple stages, so that model outputs remain aligned with the downstream tasks and decisions. The central idea is that by incorporating domain knowledge, standardized expert feedback loops, and task-relevant training strategies from the user perspective, geospatial AI workflows can be made more reliable for users. This central idea is supported by three chapters that incorporate guidance at different stages of the workflow across multiple Earth observation applications. In the first chapter, I incorporate statistical priors into generative AI models for satellite image quality enhancement. I introduced a curriculum-guided diffusion model, conditioned on a classical speckle filter, the Lee filter, that removes speckle noise from synthetic aperture radar (SAR) imagery while preserving the structures that downstream object detection requires, with ship detection improvements on SAR imagery outside the training data. In the second chapter, I standardize expert feedback protocols for transparency and downstream utility. I demonstrated a framework for expert-in-the-loop workflows that converts object detections into counts, areas, and calibrated uncertainties for lined farm ponds in Maharashtra, India, with transparency and replicability that support decision-making. The third chapter incorporates spatiotemporal and environmental context into foundation model evaluation. I test the consistency of geospatial foundation model embeddings for predicting major soil elements across the conterminous United States, which shows that performance depends on sampling density, embedding-space proximity, and geological context, and that adding geology and soil covariates improves prediction in regions where field measurements are scarce and out of the training data distribution. These three chapters show that geospatial AI workflow outputs become scientifically useful when downstream tasks, expert review, and environmental context guide both workflow design and evaluation. Reliability, when designed into multiple stages of the workflow, gives users justified confidence in the resulting measurements.