A Physics-Informed Convolutional Neural Network for Global Ocean PAR Gap-Filling From Satellite Observations
衛星観測からの全球海洋PARギャップフィリングのための物理情報に基づく畳み込みニューラルネットワーク (AI 翻訳)
Jing Tan, R. Frouin, Shuo Liu
🤖 gxceed AI 要約
日本語
海洋の光合成有効放射(PAR)の衛星データには観測ギャップが存在し、生物地球化学モデルの精度を低下させる。本研究では、畳み込みニューラルネットワーク(CNN)を用いて雲データを入力とし、物理的一貫性を保った全天候PARデータセットを生成した。MODIS-Aquaデータでは高精度(R²=0.96)を達成し、EPIC観測との比較でも物理的リアリズムを確認。本手法は海洋炭素循環の長期モニタリングやデータ同化に貢献する。
English
Satellite-derived photosynthetically available radiation (PAR) suffers from persistent data gaps. This study develops a physics-informed convolutional neural network (CNN) to produce gap-free daily PAR fields at 0.5° resolution using MODIS-Aqua and MERRA-2 cloud data. The CNN achieves high accuracy (R²=0.96) with satellite inputs and reasonable performance with reanalysis data, validated against independent EPIC estimates. The gap-filled product provides robust forcing for ocean-biogeochemistry models and climate trend analysis.
Unofficial AI-generated summary based on the public title and abstract. Not an official translation.
📝 gxceed 編集解説 — Why this matters
日本のGX文脈において
日本周辺海域の一次生産や水産資源評価にも応用可能な基礎技術。気候変動適応策の基礎データとして、海洋生態系モデルの高度化に寄与する。
In the global GX context
PAR gap-filling improves ocean carbon cycle and primary productivity modeling globally, supporting climate change research and biogeochemical data assimilation. The physics-informed CNN approach is transferable to other satellite-derived environmental variables.
👥 読者別の含意
🔬研究者:Useful for ocean biogeochemical modelers and remote sensing scientists seeking gap-free PAR fields for carbon cycle studies.
📄 Abstract(原文)
Accurate characterization of photosynthetically available radiation (PAR) is essential for quantifying marine primary production and carbon cycling. However, current satellite sensors suffer from persistent data gaps due to orbital limitations, sun glint, and large solar zenith angles. Such gaps bias bloom phenology, obscure long-term trends, and degrade biogeochemical models requiring continuous radiative forcing. To address this, we developed a deep convolutional neural network (CNN) to generate gap-free daily PAR fields at 0.5 resolution using MODIS-Aqua and MERRA-2 datasets. Unlike traditional methods that rely solely on spatiotemporal patterns or reanalysis alone, the CNN utilizes cotemporal MODIS-Aqua satellite-derived cloud fractional coverage (CC) and optical thickness, seamlessly substituted by MERRA-2 in gap regions, to account for cloud-modulated PAR variability. The model demonstrates high accuracy using MODIS-Aqua inputs [R<inline-formula> <tex-math notation="LaTeX">${}^{2}=0.96$ </tex-math></inline-formula>, <inline-formula> <tex-math notation="LaTeX">$\text {bias} = -0.19$ </tex-math></inline-formula> E/m2/day (−0.5%), root-mean-square error (RMSE) = 3.20 E/m2/day (8.3%)], though performance degrades with MERRA-2 cloud data [R<inline-formula> <tex-math notation="LaTeX">${}^{2}=0.84$ </tex-math></inline-formula>, bias = 0.17 E/m2/day (−0.4%), RMSE = 6.47 E/m2/day (16.7%)]. Comparisons with independent earth polychromatic imaging camera (EPIC) estimates (R<inline-formula> <tex-math notation="LaTeX">${}^{2}=0.82$ </tex-math></inline-formula>, RMSE = 6.43 E/m2/day, <inline-formula> <tex-math notation="LaTeX">$\text {bias} = -0.85$ </tex-math></inline-formula> E/m2/day) confirm the physical realism of the reconstructed fields. This gapless product provides a robust forcing field for ocean-biogeochemistry models and data assimilation, enabling more accurate assessments of long-term climate trends. Future work will integrate multisensor fusion and higher-resolution cloud inputs.
🔗 Provenance — このレコードを発見したソース
- semanticscholar https://doi.org/10.1109/tgrs.2026.3696266first seen 2026-06-29 09:07:24
🔔 こうした論文の新着を逃したくない方は キーワードアラート に登録(無料・3キーワードまで)。
gxceed は公開メタデータに基づく研究支援データセットです。要約・翻訳・解説は AI 支援で生成されています。 最終的な解釈・検証は利用者が原典資料に基づいて行うことを前提とします。