<b>OKG-ConvGRU: A domain knowledge-guided remote sensing prediction framework for ocean elements</b>

1.The data folder storesThe data folder stores the long time-series remote sensing image data used in the experiment, which has been preprocessed. The study area is the eastern China Sea, and we selected the chlorophyll a concentration Chl-a as the target element for model prediction, and its influencing factors include sea surface temperature SST, particulate inorganic carbon PIC, particulate organic carbon POC, photosynthetically active radiation PAR, and normalized fluorescence line brightness NFLH.<br><br>In this study, chlorophyll-a concentration (Chl-a) was selected as the target element for model prediction.Chl-a is influenced by sea surface temperature (SST), particulate inorganic carbon (PIC), particulate organic carbon (POC), photosynthetically active radiation (PAR), and normalized fluorescence line brightness (NFLH) (Zhaiet al. 2021). According to existing studies, phytoplankton growth is affected by multiple interactions of physical, chemical, and biological factors (Zhang et al. 2023; Menget al. 2022). Among these factors, SST showsa significant correlation with Chl-a concentration (Chen, Cai,et al. 2024), while interactions among POC, PIC, and Chl-a reflect the productivity and carbon cycling processes in marine ecosystems (Dong et al. 2025;Karmakaret al. 2024). In addition, PAR is strongly positively correlated with Chl-a (McGintyet al. 2016; Wang et al. 2020).The experimental data were obtained from satellite remote sensing images provided by NASA, spanning approximately 22 years from August 2002 to May 2024, with a monthlytemporal resolution. The data were derived from the MODIS L3 OceanColor product, available through anopen-access website (https://oceancolor.gsfc.nasa.gov/l3/), with a spatial resolution of 4 km.Data pre-processing：In this part, we performedseveral preprocessing operations on the original satellite images to improve the data qualityand make them better adapt to the subsequent spatio-temporal prediction. To address the issue of missing values in original images, the data interpolation empirical orthogonal function (DINEOF) method (Wang, Gao, and Liu2019; Beckers, Barth, and Alvera-Azcárate2006) was utilized to reconstruct the missing image data. This method effectively restores the missing values and retains the spatio-temporal variation characteristics of the data through spatio-temporal covariance matrix decomposition and iterative interpolation. Subsequently, high-precision land vector data corresponding to the selected projection was employed to implement a masking process for the land anomalies of the ocean water color data, thereby eliminating geographic interference. To unify the dimensions of the multi-source data, the parameters were normalized to the [0,1] interval by Min-Max normalization (Prasetyowatiet al. 2022). Finally, the images were uniformly cropped to 320×568 pixel specifications to fit the model inputs.The dataset division strictly followed the principle of temporal continuity, and the 262 months of data from August 2002 to May 2024 (2002.08-2024.05) were divided into three subsets: the training set (2002.08-2018.05, 90 months) is used for model parameter learning, the validation set (2018.06-2021.05, 36 months) is used for hyperparameter optimization, and the test set (2021.06-2024.05, 36 months) is used to evaluate the model generalization ability.2.The OKG folderThe OKG folder stores the source code of our constructed remote sensing spatio-temporal knowledge graph (OKG) of ocean elements as well as the semantic representation process, which contains the knowledge graph visualization, storage to Neo4j, and embedded models (TransE,TransH) training to evaluate the visualization process.3.The cross_convgru folderThe cross_convgru folder contains the source code of the developed model.4.Experimental environmentThe experiments are conducted on a workstation that is equipped with an Intel Core i7-14650HX processor and operates on the Windows 11 operating system. The model is implemented based on the PyTorch framework and utilizes an NVIDIA RTX 4070 graphics card (32GB video memory) for the purpose of training acceleration, with CUDA version 12.5.Code development and debugging are conducted in the PyCharm integrated development environment.<br>5.The excel folderThe excel folder stores all the tabular data used in the thesis, which contains the values of the indicators obtained from the various experiments.6.The pictures folderThe pictures folder stores all the pictures presented in the manuscript of the paper, including module flowcharts, visualized knowledge graphs, predictions of the model, etc.1. data文件夹：data文件夹存储本实验所用的预处理后长时序遥感影像数据集。研究区域为中国东海，我们选取叶绿素a浓度（Chlorophyll a concentration, Chl-a）作为模型预测的目标要素，其影响因子包括海表温度（Sea Surface Temperature, SST）、颗粒无机碳（Particulate Inorganic Carbon, PIC）、颗粒有机碳（Particulate Organic Carbon, POC）、光合有效辐射（Photosynthetically Active Radiation, PAR）以及归一化荧光线亮度（Normalized Fluorescence Line Height, NFLH）。 本研究选取叶绿素a浓度（Chl-a）作为模型预测目标。Chl-a受海表温度（SST）、颗粒无机碳（PIC）、颗粒有机碳（POC）、光合有效辐射（PAR）及归一化荧光线亮度（NFLH）调控（Zhai et al. 2021）。已有研究表明，浮游植物生长受物理、化学及生物因子的多重交互作用影响（Zhang et al. 2023; Meng et al. 2022）。其中，SST与Chl-a浓度呈显著相关关系（Chen, Cai, et al. 2024）；而POC、PIC与Chl-a之间的交互作用可反映海洋生态系统的生产力与碳循环过程（Dong et al. 2025; Karmakar et al. 2024）。此外，PAR与Chl-a呈显著正相关（McGinty et al. 2016; Wang et al. 2020）。本实验数据源自美国国家航空航天局（National Aeronautics and Space Administration, NASA）提供的卫星遥感影像，时间跨度为2002年8月至2024年5月，共计约22年，时间分辨率为月度。数据来源于MODIS L3海洋水色产品，可通过开放获取网站（https://oceancolor.gsfc.nasa.gov/l3/）获取，空间分辨率为4 km。 数据预处理：本环节对原始卫星影像开展多项预处理操作，以提升数据质量并使其更适配后续时空预测任务。针对原始影像存在的缺失值问题，我们采用数据插值经验正交函数（Data Interpolation Empirical Orthogonal Function, DINEOF）方法（Wang, Gao, and Liu 2019; Beckers, Barth, and Alvera-Azcárate 2006）对缺失的影像数据进行重构。该方法通过时空协方差矩阵分解与迭代插值，可有效修复缺失值并保留数据的时空变化特征。随后，我们采用与所选投影匹配的高精度陆地矢量数据，对海洋水色数据的陆地异常区域进行掩膜处理，以消除地理干扰。为统一多源数据的量纲，我们通过最小-最大归一化（Min-Max Normalization）方法将各参数归一化至[0,1]区间（Prasetyowati et al. 2022）。最后，将影像统一裁剪至320×568像素的规格，以适配模型输入要求。 本数据集严格遵循时空连续性原则进行划分，将2002年8月至2024年5月共计262个月的数据集划分为三个子集：训练集（2002.08-2018.05，共90个月）用于模型参数学习；验证集（2018.06-2021.05，共36个月）用于超参数优化；测试集（2021.06-2024.05，共36个月）用于评估模型的泛化能力。 2. OKG文件夹：OKG文件夹存储我们构建的海洋要素遥感时空知识图谱（Ocean Knowledge Graph, OKG）的源代码及语义表征流程，涵盖知识图谱可视化、存储至Neo4j数据库以及嵌入模型（TransE、TransH）训练与评估可视化流程。 3. cross_convgru文件夹：cross_convgru文件夹包含所开发模型的源代码。 4. 实验环境：本实验在搭载Intel Core i7-14650HX处理器、运行Windows 11操作系统的工作站上完成。模型基于PyTorch框架实现，并采用NVIDIA RTX 4070显卡（32GB显存）进行训练加速，CUDA版本为12.5。代码开发与调试工作在PyCharm集成开发环境中完成。 5. excel文件夹：excel文件夹存储本论文所用的全部表格数据，包含各项实验得到的指标数值。 6. pictures文件夹：pictures文件夹存储论文手稿中呈现的全部图片，包括模块流程图、知识图谱可视化结果、模型预测结果等。