A Unified BRDF-based Optimization Framework for the Simultaneous Inversion of Aerosol Optical Depth and Surface Reflectance from GK-2A/AMI Measurements

다국어 초록 (Multilingual Abstract)

Aerosol optical depth （AOD） retrieval from visible-channel satellite imagery requires an accurate estimate of the land surface reflectance, which most operational algorithms obtain through the shortwave-infrared （SWIR） spectral ratio. GK-2A/Advanced Meteorological Imager （AMI）, the geostationary imager operated by the Korea Meteorological Administration, lacks a SWIR channel, and its operational AOD product achieves a correlation of R = 0.59 and a root-mean-square error （RMSE） of 0.25 against the Aerosol Robotic Network （AERONET）. This dissertation develops a unified bidirectional reflectance distribution function （BRDF）- based framework that simultaneously retrieves AOD and surface reflectance in a self-consistent manner, replacing the SWIR constraint with a temporal BRDF inversion that exploits the high-cadence geostationary sampling of GK-2A. Two representative retrieval approaches （the Visible Infrared Imaging Radiometer Suite （VIIRS） Enterprise spectral relationship and the AERUS-GEO Kalman-filter-based daily inversion） were first adapted to the GK-2A channel configuration. The spectral relationship failed because the SWIR channel anchoring the surface reflectance prediction is absent; the blue-channel surface reflectance changed by up to 97.7% over the AOD range 0.1–0.8, while the red channel changed by only 14.5%, and the fixed low-AOD training assumption could not accommodate this asymmetry （R = 0.34, slope = 0.12）. The Kalman-filter approach, operating in the red channel, failed for a different reason: a Second Simulation of a Satellite Signal in the Solar Spectrum–Vector （6SV） simulation showed that the red-channel top-of- canopy （TOC） reflectance changed by only 0.016 over an AOD range of 0.1–0.4, leaving the retrieval underdetermined regardless of the covariance configuration or prior information. A 14-day sliding-window BRDF inversion was then implemented in the Ross–Thick Li-Sparse Reciprocal （RTLS） kernel basis to provide the surface constraint. A multi-year Moderate Resolution Imaging Spectroradiometer （MODIS） climatology and a daily free inversion produced opposite-sign AOD biases （−0.040 and +0.072）, and a perturbation analysis quantified the propagation: a +0.01 blue surface reflectance error shifted the retrieved AOD by −0.06 to −0.10. This opposite-sign structure motivated a Bayesian optimization strategy that combines the stability of the climatology with the adaptability of the daily inversion, enabling the framework to update the surface reflectance and AOD iteratively within each retrieval cycle. The framework crosses three BRDF strategies （free scaled-RTLS, Bayesian with MCD43D prior, and Bayesian with MCD19A3D prior） with two aerosol look-up table families （VIIRS Enterprise two-model and Multi-Angle Implementation of Atmospheric Correction （MAIAC） five-model）, yielding six configurations unified under a common retrieval architecture. A hybrid method map assigns the optimal configuration to each Köppen–Geiger climate zone, with the eight geographic regions retained for reporting. Validated against 74 AERONET sites for 2023, the hybrid retrieval achieves R = 0.906, RMSE = 0.130, and 64.6% expected-error compliance. The simultaneously produced surface reflectance reaches an RMSE of 0.0153 in the blue band and 0.0094 in the red band against an AERONET-derived reference, confirming the self-consistency of the AOD and surface reflectance retrievals. Temporal stability was assessed by applying the 2023-optimized method map unchanged to 2024; the aggregate RMSE remained at 0.130, with seven of eight regions performing at least as well as in 2023. The comparison against the National Meteorological Satellite Center （NMSC） product showed a 42–43% reduction in blue-band surface reflectance RMSE in both years. An intercomparison with the Himawari-9/AHI aerosol product on the overlapping sub-domain yielded a lower RMSE （0.128 versus 0.209） and higher expected-error compliance （64.6% versus 55.1%）. Against the polar-orbiting MODIS MAIAC products, the retrieval held the same aggregate RMSE （0.132） as the reference on a common matchup set. Three regions （Taiwan, mainland Southeast Asia, and South Asia） remain below the 67% validation threshold in both years, pointing to residual difficulties in cloud screening and aerosol typing shared by current geostationary retrievals. The framework does not depend on a SWIR channel, and its requirements （blue and red visible channels, a sub-hourly revisit, and a BRDF climatology） are met by other current geostationary imagers.

목차 (Table of Contents)

• CHAPTER 1. Introduction 2
• 1.1. Background 2
• 1.2. Motivation and Problems 11
• 1.3. Objectives and Scope of This Dissertation 18
• CHAPTER 2. Study Area and Data 30
• 2.1. GK-2A/AMI 32
• 2.2. Radiative Transfer Lookup Table 35
• 2.3. Aerosol Optical Models 37
• 2.4. MODIS BRDF Products 40
• 2.5. Ground-based Validation Data 42
• 2.6. Auxiliary Data 47
• 2.7. Study Area 50
• CHAPTER 3. Application and Analysis of Existing AOD Retrieval Algorithms 53
• 3.1. Introduction 53
• 3.2. Methods 55
• 3.3. Results 64
• 3.4. Summary 69
• CHAPTER 4. Feasibility of AOD Retrieval Using Existing BRDF Resources 72
• 4.1. Introduction 72
• 4.2. Methods 75
• 4.3. Results 81
• 4.4. Summary and Implications 102
• CHAPTER 5. Retrieval Framework Design and Validation 106
• 5.1. Introduction 106
• 5.2. Methods 117
• 5.3. Results 126
• 5.4. Summary 147
• CHAPTER 6. Operational Assessment and Intercomparison 151
• 6.1. Introduction 151
• 6.2. Multi-Year Performance: 2024 Extension 154
• 6.3. Intercomparison with the NMSC Operational Product 160
• 6.4. Intercomparison with Himawari-9/AHI 177
• 6.5. Comparison with MODIS MAIAC Products 183
• 6.6. Summary 187
• CHAPTER 7. General Conclusion 192
• 7.1. Conclusion 192
• 7.2. Contributions 197
• 7.3. Directions for Future Research 200