OKG-ConvGRU: A domain knowledge-guided remote sensing prediction framework for ocean elements

Abstract: Accurately predicting key elements of the ocean, such as chlorophyll-a concentration and sea surface temperature, is crucial for maintaining marine ecological balance, responding to marine disasters and pollution, and promoting the sustainable use of marine resources. The existing spatiotemporal prediction models mainly rely on physical or data-driven methods. Physical models are limited by modeling complexity and parameterization errors, while data-driven models lack interpretability and rely on high-quality data. To address these challenges, this study proposes OKG ConvGRU, a domain knowledge guided ocean element remote sensing prediction framework. This framework integrates knowledge graph with ConvGRU network, utilizing prior knowledge of ocean science to enhance the predictive performance of ocean elements in remote sensing images. Firstly, we constructed a spatiotemporal knowledge graph (OKG) of ocean elements, and then performed semantic embedding representations on their spatial and temporal dimensions. Subsequently, a cross attention based feature fusion module (CAFM) was designed to effectively integrate spatiotemporal multimodal features. Finally, these fusion functions are integrated into the enhanced ConvGRU network. For multi-step prediction, we adopt the Seq2Seq architecture combined with a multi-step rolling strategy. The prediction experiment of chlorophyll-a concentration in the eastern waters of China verified the effectiveness of the proposed framework. The results indicate that compared with the baseline model, OKG ConvGRU exhibits significant advantages in prediction accuracy, long-term stability, data utilization efficiency, and robustness. This study provides scientific basis and technical support for precise monitoring and sustainable development of marine ecological environment. File Description: The data folder stores the long time series remote sensing image data used in the experiment, which has been preprocessed. The research area is the eastern sea area, and chlorophyll a concentration Chl-a is selected as the target element for model prediction. The 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. The cross_comvgru folder contains the source code of developed models. The OKG folder stores the source code and semantic representation process of our constructed ocean element remote sensing spatiotemporal knowledge graph (OKG), which includes knowledge graph visualization, storage to Neo4j, and embedded model (TransE, TransH) training to evaluate the visualization process. Experimental environment: The experiment was conducted on a workstation equipped with an Intel Core i7-14650HX processor and running on a Windows 11 system. This model is implemented based on the PyTorch framework and utilizes NVIDIA RTX 4070 graphics card (32GB of video memory) for training acceleration, CUDA version 12.5.