Atmospheric Correction for HY-3A CZI: A Deep Learning Method Applied Across Open Ocean, Coastal, and Turbid Inland Waters

Atmospheric correction (AC) is a critical step for retrieving accurate remote sensing reflectance (Rrs). However, in highly turbid inland and coastal waters, traditional AC algorithms typically require shortwave infrared (SWIR) or ultraviolet (UV) bands, thereby limiting their applicability to sensors that lack such bands (e.g., HY-3A CZI). To address these limitations, we introduce ACTransNet, an end-to-end deep learning model based on the Transformer architecture. This model directly retrieves Rrs from satellite digital numbers (DNs), bypassing the explicit radiometric calibration (RC) process. The model is trained on synthetic datasets derived from Sentinel-2 MSI, with reference Rrs generated using a combined strategy that involves the POLYMER and ACOLITE-DSF algorithms. The training data span 40 distinct regions, covering diverse water environments ranging from clear open ocean to turbid coastal and inland waters. These data undergo a strict quality control process to ensure their representativeness and reliability. Application results using CZI data demonstrate that ACTransNet performs robustly across seven spectral bands, withR2values ranging from 0.61 to 0.97. In the visible bands, ACTransNet substantially reduces the mean absolute percentage difference (MAPD), with an average reduction of ∼ 82% relative to ACOLITE-DSF and ∼ 74% relative to NIR-bio. Comparisons with quasi-synchronized Sentinel-2 MSI imagery further confirm that the model maintains strong spatial consistency (R2> 0.96) across two contrasting water environments (the eastern coast of the Bay of Bengal and Lake Hulun). In a separate evaluation, the model also outperforms traditional multi-step AC algorithms, such as DSF and NIR-bio. Overall, ACTransNet integrates the strengths of both POLYMER and DSF, offering enhanced accuracy and robustness across diverse water environments. This model provides a practical and generalizable solution for AC tasks, particularly for sensors with limited spectral bands, and offers valuable references for remote sensing applications on other similar satellite platforms.