Date of Award

2026-08-01

Degree Name

Doctor of Philosophy

Department

Mathematical Sciences

Advisor(s)

Maria C. Mariani

Abstract

Passive microwave (PMW) radiometers are a primary source of global satellite precipitation estimates, but their retrievals are degraded by parallax-induced geolocation displacement, orbit-level artifacts, and environment-dependent biases. This dissertation develops physically based and machine learning-driven methods to assess the quality and increase the reliability of PMW precipitation retrievals, using the Global Precipitation Measurement (GPM) mission and its Goddard Profiling Algorithm (GPROF) as the primary testbed. It comprises three independent but complementary studies centered on the diagnosis of retrieval errors and the detection of observational artifacts. The first study introduces a physically based parallax correction that realigns cloud-level microwave signals with their true surface locations using freezing-level information, thereby improving retrieval accuracy over the conterminous United States, particularly during the summer season. The second develops a sensor-adaptive, incremental-learning computer vision framework that detects and classifies artifacts in global PMW rainfall products from scarce labeled examples, matching operational detectors performance while adding interpretability and the ability to adapt as new sensors are added. The third presents a machine learning framework that diagnoses average instantaneous GPROF bias across tropical regions using atmospheric context rather than a fixed ground reference, showing that reanalysis-derived predictors can be used to recognize and diagnose PMW biases. Together, the three studies show that pairing physical reasoning with machine learning advances the diagnosis and detection of PMW retrieval limitations. The latter two chapters also demonstrate methods that remain effective where dense ground validation is unavailable, which is usually the situation across much of the globe.

Language

en

Provenance

Received from ProQuest

File Size

107 p.

File Format

application/pdf

Rights Holder

Andres Felipe Monsalve Salazar

Download

Share

COinS