Satellite remote sensing plays a fundamental role in observing oceanic processes by providing large-scale, long-term, and continuous measurements. With the increasing availability of multisource satellite data, challenges such as data gaps, complex environmental conditions, and the limitations of conventional retrieval methods have become more evident. In recent years, artificial intelligence (AI) has emerged as a practical and effective approach to address these issues. This article reviews the development of AI techniques in satellite ocean remote sensing, focusing on three main application areas: parameter retrieval, data reconstruction, and image-based ocean phenomenon detection. For geophysical variable retrieval, AI models such as convolutional neural networks (CNNs) and Transformer architectures have improved the accuracy of ocean waves, sea surface, salinity, wind, and ocean color estimates, especially under extreme or noisy conditions. In the field of data reconstruction, AI methods enable the completion of missing data in both surface and subsurface ocean layers, offering finer spatial–temporal resolution and better consistency than traditional interpolation approaches. For image interpretation, deep learning (DL) models have been applied to detect and segment dynamic ocean features such as mesoscale eddies, internal waves, sea ice, and tropical cyclones (TCs), achieving high efficiency and precision. This article also highlights the integration of AI with physical knowledge, the use of multisource fusion, and the trend toward near real-time (NRT) applications. These developments indicate that AI will play an increasingly important role in future satellite-based ocean observation and environmental monitoring.

During the melting season, predicting the daily sea ice concentration (SIC) of the Pan-Arctic at a subseasonal scale is strongly required for economic activities and a challenging task for current studies. We propose a deep-learning-based data-driven model to predict the 90 days SIC of the Pan-Arctic, named SICNet90. SICNet90takes the historical 60 days’ SIC and its anomaly and outputs the SIC of the next 90 days. We design a physically constrained loss function, normalized integrated ice-edge error (NIIEE), to constrain the SICNet$_{\mathrm {90{'}s}}$optimization by the spatial morphology of SIC. The satellite-observed SIC trains (1991–2011/1997–2017) and tests the model (2012/2018–2020). For each test year, a 90-day SIC prediction is made daily from May 1 to July 2. The binary accuracy (BACC) of sea ice extent (SIC$>$15%) and the mean absolute error (MAE) are evaluation metrics. Experiments show that SICNet90significantly outperforms the Climatology benchmark on 90 days prediction, with a BACC/MAE improvement/reduction of 5.41%/1.35%. The data-driven model shows a late-spring-early-summer predictability barrier (around June 20) and a prediction challenge (around July 10), consistent with SIC’s autocorrelation. The NIIEE loss optimizes the predictability barrier/challenge with a BACC increase of 4%. Using a 60 days historical SIC to predict 90 days SIC is better than a historical SIC of 30/90 days. The historical 2-m surface air temperature shows positive contributions to the prediction made from May 1 to mid-June, but negative contributions to the prediction made after mid-June. The historical sea surface temperature and 500 hp geopotential height show negative contributions.

This study develops a deep learning (DL) model to classify the sea ice and open water from synthetic aperture radar (SAR) images. We use the U-Net, a well-known fully convolutional network (FCN) for pixel-level segmentation, as the model backbone. We employ a DL-based feature extracting model, ResNet-34, as the encoder of the U-Net. To achieve high accuracy classifications, we integrate the dual-attention mechanism into the original U-Net to improve the feature representations, forming a dual-attention U-Net model (DAU-Net). The SAR images are obtained from Sentinel-1A. The dual-polarized information and the incident angle of SAR images are model inputs. We used 15 dual-polarized images acquired near the Bering Sea to train the model and employ the other three images to test the model. Experiments show that the DAU-Net could achieve pixel-level classification; the dual-attention mechanism can improve the classification accuracy. Compared with the original U-Net, DAU-Net improves the intersection over union (IoU) by 7.48.% points, 0.96.% points, and 0.83.% points on three test images. Compared with the recently published model DenseNetFCN, the three improvement IoU values of DAU-Net are 3.04.% points, 2.53.% points, and 2.26.% points, respectively.

This study develops a deep learning (DL) model to extract the ship size from Sentinel-1 synthetic aperture radar (SAR) images, named SSENet. We employ a single shot multibox detector (SSD)-based model to generate a rotatable bounding box (RBB) for the ship. We design a deep-neural-network (DNN)-based regression model to estimate the accurate ship size. The hybrid inputs to the DNN-based model include the initial ship size and orientation angle obtained from the RBB and the abstracted features extracted from the input SAR image. We design a custom loss function named mean scaled square error (MSSE) to optimize the DNN-based model. The DNN-based model is concatenated with the SSD-based model to form the integrated SSENet. We employ a subset of the OpenSARShip, a data set dedicated to Sentinel-1 ship interpretation, to train and test SSENet. The training/testing data set includes 1500/390 ship samples. Experiments show that SSENet is capable of extracting the ship size from SAR images end to end. The mean absolute errors (MAEs) are under 0.8 pixels, and their length and width are 7.88 and 2.23 m, respectively. The hybrid input significantly improves the model performance. The MSSE reduces the MAE of length by nearly 1 m and increases the MAE of width by 0.03m compared to the mean square error (MSE) loss function. Compared with the well-performed gradient boosting regression (GBR) model, SSENet reduces the MAE of length by nearly 2 m (18.68%) and that of width by 0.06 m (2.51%). SSENet shows robustness on different training/testing sets.

This study proposes a purely data-driven model for the weekly prediction of daily sea ice concentration (SIC) of the pan-Arctic (90 N, 45 N, 180 E, 180 W) during the melting season. The model, SICNet, adopts an encoder–decoder framework with fully convolutional networks (FCNs) and can predict the SIC (covering$320\times224$grids, each with a resolution of 25 km) one-week lead with high accuracy. We design a temporal–spatial attention module (TSAM) to help SICNet capture spatiotemporal dependencies from SIC sequences. The satellite-derived SIC data of 33 years (1988–2020) from the National Snow and Ice Data Center (NSIDC) are employed to train and test the model, 1988–2015 for training, and 2016–2020 for testing. SICNet achieves the mean absolute error (MAE) of 2.67%, the mean absolute percentage error (MAPE) of 8.67%, and the Nash–Sutcliffe efficiency (NSE) of 0.9784 in weekly predicting of SIC during the melting season. SICNet achieves better performance than existing deep-learning-based models. The TSAM reduced the MAE from 2.73% to 2.67%. We evaluate the model’s performance by recursively predicting, from seven- to 28-day leads. We employ the binary accuracy (BACC) metric to measure the accuracy of the predicted sea ice extent (SIE) and compare SICNet with the anomaly persistence (Persist). SICNet shows better performance than Persist with an average BACC on the 28th day of 2016–2019 over 90% (90.17%). For the 28-day lead predictions of three extreme minimum SIE in September 2007, 2012, and 2020, SICNet outperforms Persist with an average improvement of 1.84% in BACC and$0.16 milkm^{2}$in the SIE error.

Sea ice type's classification plays an essential role in Arctic marine navigation. Synthetic Aperture Radar (SAR) is independent of weather conditions and is widely used in sea ice classification. U-Net is a well-performed deep learning (DL) framework in image classification. This study constructs a U-Net-based “end-to-end” model to classify multi-year ice (MYI), first-year ice (FYI), and open water. We label the SAR images by ice chart provided by the U.S. National Ice Center (USNIC). The labeled images are divided into chips to be fed into the U-Net model for training and testing. Experiments show that the precision and the recall of the testing set are 89.55% and 89.46%. The proposed model can classify sea ice types in an “end-to-end” way with high accuracy.

This study proposes a fully-convolutional-networks-based (FCN-based) model to predict the daily Arctic SIC in the melt season. The FCN-based model adopts an encoder-decoder framework with FCN layers as the basic unit. We use the SIC series of the last seven days to predict the SIC of the coming day. The Arctic SIC series of 31 years (1988–2018) from the NSIDC are employed to train and evaluate the model. Experiments show that the FCN-based model is capable of accurately predicting the daily Arctic SIC of all grids (320× 224) in an end-to-end way, with an MAE under 1%. The FCN-based model shows apparent advantages over the newly published convolution-neural-network-based (CNN-based) model in prediction accuracy, efficiency, and resource occupation.

Mesoscale eddies are circular flowing currents that can retain and transport salt, heat, and nutrients all around the ocean. Mesoscale eddies can be detected on remote sensing images, i.e., sea surface height (SSH) images, sea surface temperature (SST) images, chlorophyll concentration images, etc. Most existing automatic eddy detection algorithms are developed based on one kind of remote sensing data. There is a lack of an automatic eddy detection algorithm that can make full use of multi-source remote sensing data to ensure the accuracy and efficiency of eddy detection. The paper proposes a multi-modal U-Net model, a deep neural network-based framework for eddy detection from multi-source remote sensing images. Compared with the previous eddy detection methods, the newly proposed method improves the accuracy and efficiency of eddy detection by using fusion data of SSH and SST.

Accurate and robust classification methods of sea ice and open water are significant for many applications. Synthetic Aperture Radar (SAR) imaging capability is independent of weather conditions and is widely used in sea ice classification. U-Net, a deep learning framework, has achieved great success in the field of biomedical image classification. In this study, we construct a U-Net-based “end-to-end” model to classify the sea ice and open water pixels in SAR imagery. Five SAR images acquired in the Gulf of Alaska near Bering Strait are used in this case study. We manually label the SAR images as ice and water. The labeled images from the first four SAR image are divided into chips to be fed into the U-Net model for training. The fifth SAR image is employed as the testing data. Experiments show that the precision and the recall of the testing image is 91.64% and 91.70%, respectively. Most of the sea ice, including small chunks and sinuous ice edges, can be successfully classified.

The spatio-temporal residual network (ST-ResNet) leverages the power of deep learning (DL) for predicting the volume of citywide spatio-temporal flows. However, this model, neglects the dynamic dependency of the input flows in the temporal dimension, which affects what spatio-temporal features may be captured in the result. This study introduces a long short-term memory (LSTM) neural network into the ST-ResNet to form a hybrid integrated-DL model to predict the volumes of citywide spatio-temporal flows (called HIDLST). The new model can dynamically learn the temporal dependency among flows via the feedback connection in the LSTM to improve accurate captures of spatio-temporal features in the flows. We test the HIDLST model by predicting the volumes of citywide taxi flows in Beijing, China. We tune the hyperparameters of the HIDLST model to optimize the prediction accuracy. A comparative study shows that the proposed model consistently outperforms ST-ResNet and several other typical DL-based models on prediction accuracy. Furthermore, we discuss the distribution of prediction errors and the contributions of the different spatio-temporal patterns.

Short-term traffic forecasting on large street networks is significant in transportation and urban management, such as real-time route guidance and congestion alleviation. Nevertheless, it is very challenging to obtain high prediction accuracy with reasonable computational cost due to the complex spatial dependency on the traffic network and the time-varying traffic patterns. To address these issues, this paper develops a residual graph convolution long short-term memory (RGC-LSTM) model for spatial-temporal data forecasting considering the network topology. This model integrates a new graph convolution operator for spatial modelling on networks and a residual LSTM structure for temporal modelling considering multiple periodicities. The proposed model has few parameters, low computational complexity, and a fast convergence rate. The framework is evaluated on both the 10-min traffic speed data from Shanghai, China and the 5-min Caltrans Performance Measurement System (PeMS) traffic flow data. Experiments show the advantages of the proposed approach over various state-of-the-art baselines, as well as consistent performance across different datasets.

Recently, researchers have introduced deep learning methods such as convolutional neural networks (CNN) to model spatio-temporal data and achieved better results than those with conventional methods. However, these CNN-based models employ a grid map to represent spatial data, which is unsuitable for road-network-based data. To address this problem, we propose a deep spatio-temporal residual neural network for road-network-based data modeling (DSTR-RNet). The proposed model constructs locally-connected neural network layers (LCNR) to model road network topology and integrates residual learning to model the spatio-temporal dependency. We test the DSTR-RNet by predicting the traffic flow of Didi cab service, in an 8-km2 region with 2,616 road segments in Chengdu, China. The results demonstrate that the DSTR-RNet maintains the spatial precision and topology of the road network as well as improves the prediction accuracy. We discuss the prediction errors and compare the prediction results to those of grid-based CNN models. We also explore the sensitivity of the model to its parameters; this will aid the application of this model to network-based data modeling.