Spatiotemporal Prediction of Wind Fields in Coastal Urban Environments Using Multi-Source Satellite Data: A GeoAI Approach

Highlights

What are the main findings?

• A novel lightweight GeoAI framework (DA-DSC-UNet) is proposed to fuse multi-source satellite data (ASCAT, FY-3E, QuickSCAT) for high-resolution coastal wind field prediction.
• The model significantly outperforms existing methods by reducing the Mean Absolute Error (MAE) by 14–25.8% and demonstrates exceptional robustness against observational noise.

What are the implications of the main findings?

• The accurate wind field reconstruction provides critical decision support for urban disaster mitigation, pollutant dispersion monitoring, and infrastructure resilience in complex coastal zones.
• Due to its low computational cost, the framework enables real-time deployment on edge devices, directly contributing to the monitoring of Sustainable Development Goal 11 (Sustainable Cities).

Abstract

Rapid urbanization in coastal regions presents complex challenges for environmental management and public safety. Accurate, high-resolution wind field monitoring is critical for urban disaster mitigation, infrastructure resilience, and pollutant dispersion analysis in these densely populated areas. However, utilizing massive multi-source satellite remote sensing data for precise prediction remains difficult due to the spatiotemporal heterogeneity caused by the land–sea interface. To address this, this study proposes a novel lightweight Geospatial Artificial Intelligence (GeoAI) framework (DA-DSC-UNet) designed to predict wind fields in coastal urban environments (e.g., Fujian, China). We constructed a dataset by integrating multi-source satellite scatterometer products (including Advanced Scatterometer (ASCAT), Fengyun-3E (FY-3E), and Quick Scatterometer (QuickSCAT)) and buoy observations. The framework employs a UNet architecture enhanced with dual attention mechanisms (Efficient Channel Attention (ECA) and Convolutional Block Attention Module (CBAM)) to adaptively extract features from remote sensing signals, focusing on critical spatial regions like urban coastlines. Additionally, depthwise separable convolutions (DSCs) are introduced to ensure the model is lightweight and efficient for potential deployment in urban monitoring systems. Results demonstrate that our approach significantly outperforms existing deep learning models (reducing Mean Absolute Error (MAE) by 14–25.8%) and exhibits exceptional robustness against observational noise. This work demonstrates the potential of deep learning in enhancing the value of remote sensing data for urban resilience, sustainable development (SDG 11), and environmental monitoring in complex coastal zones.

1. Introduction

The rapid acceleration of global urbanization has concentrated population and infrastructure in coastal regions, creating complex “land-ocean-city” systems [1]. In these densely populated areas, wind field dynamics play a critical role in urban environmental management and public safety [2,3]. Accurate, high-resolution monitoring of coastal wind fields is not only essential for assessing wind energy resources [4] but is also vital for analyzing the dispersion of urban air pollutants [5], ensuring the safety of high-rise infrastructure against wind loads [6], and mitigating risks from extreme weather events such as typhoons and storm surges [7]. However, the complex thermodynamic interactions at the land–sea interface, coupled with the roughness of urban terrain, introduce significant spatiotemporal heterogeneity to wind fields, posing a severe challenge for precise environmental monitoring [8].

Traditionally, wind field data rely on ground observation stations and buoys. While accurate, these point-based observations are sparsely distributed and expensive to maintain, leaving vast observational blind spots and limiting their spatial representativeness in coastal waters and complex urban fringes [9]. Satellite remote sensing technology offers a compelling solution by providing synoptic, repetitive, and objective observations over large scales [10]. Microwave scatterometers, such as the Advanced Scatterometer (ASCAT) [11] and the preceding Quick Scatterometer (QuickSCAT) [12], and meteorological satellites (e.g., Fengyun-3E (FY-3E)) have become indispensable tools for acquiring sea surface wind vectors. Nevertheless, raw satellite products often suffer from limitations such as low temporal resolution (due to revisit cycles) and data gaps caused by orbital tracks or cloud contamination, which hinders their direct application in real-time urban monitoring [13].

To bridge these gaps, Numerical Weather Prediction (NWP) models have been widely used to simulate atmospheric states. Based on atmospheric dynamical equations, NWP effectively captures the physical evolution of large-scale wind fields [14]. Previous studies have made significant progress in this area: Li et al. [15] enhanced wind power simulation accuracy in the Baltic Sea region by optimizing Weather Research and Forecasting (WRF) model configurations; Santos-Alamillos et al. [16] improved regional wind speed simulations using high-resolution land-use data, demonstrating the significant impact of surface resolution on wind speed biases; and Pan et al. [17] integrated multi-source observations using real-time four-dimensional data assimilation (RTFDDA), effectively enhancing model timeliness and adaptability. Despite possessing relatively complete physical mechanisms, NWP models suffer from high computational costs and poor terrain adaptability. Recent evaluations indicate that NWP products often exhibit systematic biases in coastal transition zones due to the misrepresentation of sub-grid scale surface roughness [18], limiting their application in intelligent urban systems.

As a complementary approach, Geospatial Artificial Intelligence (GeoAI) has emerged as a paradigm shift, utilizing data-driven methods to learn spatiotemporal patterns from historical observations by bypassing complex physical equations [19]. Shallow machine learning models demonstrate strong nonlinear modeling capabilities: Ho et al. [20] enhanced short-term wind speed forecasting by integrating multi-source features using a random forest model; Renani et al. [21] combined feature selection with neural fuzzy inference systems; and Demetriou et al. [22] achieved real-time prediction of coastal wave height by integrating meteorological data. However, such models predominantly focus on single-point forecasting.

In recent years, deep spatiotemporal neural networks have demonstrated significant advantages. ConvLSTM [23] effectively integrates spatial and temporal features; Chen et al. [24] enhanced typhoon prediction using a hybrid 3D CNN-LSTM architecture; and Vallileka et al. [25] achieved high-precision weather field forecasting. More recently, attention has shifted towards advanced architectures to capture longer-range dependencies: Qiao et al. [26] applied Transformer-based models for global sea surface wind speed retrieval, while Daenens et al. [27] utilized Graph Neural Networks (GNNs) to handle the irregular spatial correlations for offshore wind power prediction. Furthermore, recent studies have explored deep learning frameworks for fusing multi-source satellite data to reconstruct high-resolution wind fields [28]. Compared to physical models, these GeoAI models demonstrate superior capabilities in extracting high-dimensional features from remote sensing imagery.

To further enhance the model’s ability to capture spatiotemporal characteristics, various advanced techniques have been introduced. For instance, channel attention mechanisms (e.g., Squeeze-and-Excitation (SE) [29], Efficient Channel Attention (ECA) [30]) and spatial attention mechanisms (e.g., Convolutional Block Attention Module (CBAM) [31]) are employed to amplify key features, while multi-scale architectures (e.g., Atrous Spatial Pyramid Pooling (ASPP) [32], Feature Pyramid Network (FPN) [33]) are utilized to expand the receptive field. Additionally, methods such as terrain-adaptive parameterization [34], multi-scale fusion CNNs [35], and physically informed neural networks (PINNs) [36] have improved prediction accuracy. However, existing architectures still face limitations when applied to coastal urban environments. They often treat spatial and channel features equally, failing to adaptively focus on critical geographical regions, such as land–sea boundaries and urban agglomerations. Furthermore, advanced fusion schemes generally suffer from complex model structures and high parameter costs, restricting their deployment in resource-constrained marine observation or edge computing environments.

To address these challenges, this study proposes a novel lightweight GeoAI framework (DA-DSC-UNet) for spatiotemporal wind field prediction in coastal urban areas. Leveraging multi-source satellite remote sensing data (including ASCAT, FY-3E, and QuickSCAT) and buoy observations from Fujian, China, we construct a high-resolution dataset. The proposed framework integrates a UNet backbone with dual attention mechanisms (ECA and CBAM) to enhance feature extraction from sparse remote sensing signals, and employs an Atrous Spatial Pyramid Pooling (ASPP) module to capture multi-scale atmospheric processes. This study aims to demonstrate how advanced deep learning techniques can unlock the potential of satellite data for urban environmental monitoring, providing reliable decision support for coastal city resilience.

2. Materials and Methods

To accurately monitor and predict the complex wind field dynamics impacting coastal urban environments, this study proposes a novel GeoAI framework, termed DA-DSC-UNet. Unlike traditional numerical models that often struggle with the high-resolution spatiotemporal heterogeneity inherent in urban–ocean transition zones, our data-driven approach leverages deep learning to capture intricate atmospheric processes.

As illustrated in Figure 1, the research workflow integrates multi-source remote sensing data with advanced deep learning techniques to establish a robust urban environmental monitoring system. The framework consists of four streamlined phases:

• Multi-source Data Fusion: Integrating large-scale satellite-derived wind field inversions with precise in situ buoy observations to construct a comprehensive spatiotemporal dataset covering the coastal urban agglomeration.
• GeoAI-driven Spatiotemporal Modeling: Deploying the DA-DSC-UNet, which incorporates dual-attention mechanisms and depthwise separable convolutions (DSC), to efficiently learn the nonlinear evolution of wind fields over complex terrain.
• Accuracy Assessment: Rigorous validation using independent in situ data to quantify the model’s performance in capturing local climate dynamics.
• Robustness Analysis: Evaluating the model’s stability under observational noise to ensure its reliability for practical urban environmental monitoring and disaster early warning.

Figure 1. Overall research framework of the proposed DA-DSC-UNet model. The workflow consists of four major modules: data preprocessing, model training and inference, validation and evaluation, and robustness evaluation. The process forms a closed-loop system from multi-source data integration to model optimization and stability verification.

Figure 1. Overall research framework of the proposed DA-DSC-UNet model. The workflow consists of four major modules: data preprocessing, model training and inference, validation and evaluation, and robustness evaluation. The process forms a closed-loop system from multi-source data integration to model optimization and stability verification.

2.1. Study Area and Data

The research area focuses on the coastal zone of Fujian Province, China, covering the geographic coordinates from 116°E to 121°E and 23°N to 28°N. This region is characterized by complex topography, including a jagged coastline and numerous islands, making it a challenging environment for wind field modeling. Geographic data were sourced from the Global Self-consistent, Hierarchical, High-resolution Geography Database (GSHHS). We selected the L1-level resolution data, which provides sufficient detail for visualizing the land–sea interface in this regional study. The primary dataset consists of sea surface wind inversion products provided by the Fujian Provincial Meteorological Bureau. To ensure comprehensive spatiotemporal coverage, this dataset integrates observations from multiple satellite platforms, including the Advanced Scatterometer (ASCAT) on MetOp-A/B/C, QuickSCAT, the China-France Oceanography Satellite (CFOSAT), and the FY-3E meteorological satellite. Detailed statistics of the satellite sources are presented in

Table 1. Partial inversion data of sea surface wind in Fujian Province.

Index	Lat	Lon	Ws	Freq	ASCAT-A	ASCAT-B	ASCAT-C	CFO	Quick SCAT	FY-3E	Mean
0	24.500	118.125	12.81	1	0	0	1	0	0	0	12.81
1	24.500	118.250	12.62	1	0	0	1	0	0	0	12.62
2	24.500	118.375	12.29	1	0	0	1	0	0	0	12.29
3	24.500	118.500	12.39	1	1	0	1	0	0	0	12.39
4	24.500	118.625	12.43	1	1	0	1	0	0	0	12.43
…

Lat: Latitude; Lon: Longitude; Ws: Wind Speed; Freq: Frequency; ASCAT: Advanced Scatterometer (A/B/C refer to instruments on MetOp-A/B/C satellites); CFO: China-France Oceanography Satellite; Quick SCAT: Quick Scatterometer; FY-3E: Fengyun-3E.

. In addition, real-time wind speed data from offshore buoys were used as ground truth for validation, as shown in

Table 2. Partial buoy data for surface winds in Fujian Province.

Date	Lat (°)	Lon (°)	10 min Avg. Wind Speed (m/s)	Altitude (m)
1 January 2021	24.48	118.20	5.9	10.0
2 January 2021	24.48	118.20	5.5	10.0
3 January 2021	24.48	118.20	5.4	10.0
4 January 2021	24.48	118.20	2.1	10.0
5 January 2021	24.48	118.20	4.9	10.0
…

.

To construct a high-quality dataset for deep learning, we performed the following preprocessing steps:

• Spatiotemporal Alignment: The raw observational data were spatially cropped to the study area and re-gridded with a step size of 0.125°, mapping all multi-source observations onto a regular grid. Although this spatial resolution (approximately 14 km) is relatively coarse for resolving micro-scale urban features such as street canyons, it is highly appropriate for monitoring the mesoscale background wind field covering the entire coastal urban agglomeration. This scale effectively captures regional atmospheric transport patterns and synoptic weather systems (e.g., typhoons) that drive local urban ventilation, serving as a critical boundary condition for finer-scale urban climate models.
• Quality Control: We applied threshold-based filtering to remove physical outliers. Wind speed values outside the range of 0.2 m/s to 30 m/s were discarded. Furthermore, to ensure data integrity, any local sample patch ($5\times 5$ grid) containing more than 12% missing values (i.e., more than 3 invalid points) was excluded.
• Sequence Construction: The valid data were organized into spatiotemporal sequences. We used a sliding window approach to construct input samples, where each sample consists of a sequence of 12 historical frames to predict the wind speed values for the next 12 time steps. To strictly prevent data leakage, the sliding window operates with a temporal stride of 1 time step exclusively within the respective training or validation periods. The dataset partitioning is based on a strict temporal split (Training: 2021; Validation: 2022), ensuring that there is no temporal overlap between the samples used for optimization and those used for evaluation. This design forces the model to learn the transferable spatiotemporal evolution rules rather than memorizing static spatial patterns. The final dataset covers the period from 1 January 2021 to February 2022.

2.2. GeoAI Model Architecture: DA-DSC-UNet

To effectively model the nonlinear spatiotemporal dynamics of wind fields over complex coastal urban terrain, we constructed a deep learning framework named DA-DSC-UNet. Based on the classic Encoder–Decoder architecture, this model is specifically optimized for meteorological remote sensing data through three strategic improvements designed to handle spatial heterogeneity and multi-scale dependency:

• Dual-Attention Mechanism: Embedded within the convolutional blocks to adaptively highlight geographically salient regions (e.g., coastlines, islands) and critical feature channels.
• Depthwise Separable Convolution: Replaces standard convolutions to minimize computational redundancy, facilitating efficient deployment in resource-constrained monitoring systems.
• Atrous Spatial Pyramid Pooling: Integrated at the bottleneck layer to capture multi-scale atmospheric contextual information, ranging from local urban roughness to regional circulation patterns.

The detailed design of each component is described below.

2.2.1. Dual-Attention Module for Spatiotemporal Feature Refinement

Attention mechanisms have garnered significant attention in the field of deep learning in recent years. Their core principle lies in enhancing the network’s ability to express critical information by adaptively adjusting the weights of features across different channels. To balance channel dependency and spatial saliency while maintaining lightweight architecture, this paper introduces a dual-branch attention module within the backbone blocks of UNet. This module consists of parallel ECA and Spatial CBAM, with coordinated re-weighting achieved through gated fusion at the branch terminals. Given input features $\mathbf{X}\in {\mathbb{R}}^{B\times C\times H\times W}$, where B is the batch size, C is the number of channels, and H and W are the spatial height and width, this module first generates weights in the orthogonal dimensions of channels and space. It then performs element-wise re-calibration on the original features, ultimately outputting the fused and residual-reconstructed features Y. In the channel branch, we employ ECA to capture local channel dependencies, as illustrated in Figure 2.

Specifically, global average pooling is performed along the spatial dimension of the input features to obtain the channel descriptor vector.

$$\mathbf{z}=\mathrm{GAP}\left(\mathbf{X}\right)\in {\mathbb{R}}^{B\times C},z_{b,c}={\displaystyle \frac{1}{HW}}\sum _{h=1}^H\sum _{w=1}^W{X}_{b,c,u,v}$$

(1)

The subscript b denotes the sample index, c denotes the channel index, and $(u,v)$ represents the spatial coordinates. Subsequently, a one-dimensional convolution is applied for local interaction along the channel axis:

$${\mathbf{a}}^{\left(c\right)}=\sigma \left(\mathrm{Conv}1\mathrm{D}(\mathbf{z},k)\right)\in {\mathbb{R}}^{B\times C}$$

(2)

where $\sigma (·)$ is a Sigmoid function, and the kernel size k is adaptively set based on the number of channels (typically an odd number) to model dependencies between adjacent channels. ${\mathbf{a}}^{\left(c\right)}$ is broadcasted back to a 4D tensor and element-wise multiplied with the original features to obtain the channel-recalibrated features:

$$\overline{\mathcal{X}}=X\odot \mathrm{Broadcast}\left({\mathbf{a}}^{\left(c\right)}\right)\in {\mathbb{R}}^{B\times C\times H\times W}$$

(3)

where ⊙ denotes the Hadamard matrix multiplication. In the spatial branch, we adopt the spatial attention concept from CBAM, with the workflow illustrated in Figure 3.

Encode complementary spatial responses using both average and max pooling methods. First perform average and max pooling along the channel dimension on X:

$${\mathbf{M}}_{\mathrm{avg}}={\mathrm{AvgPool}}_c\left(\mathcal{X}\right)\in {\mathbb{R}}^{B\times 1\times H\times W},{\mathbf{M}}_{\max }={\mathrm{MaxPool}}_c\left(\mathcal{X}\right)\in {\mathbb{R}}^{B\times 1\times H\times W}$$

(4)

Then, $[{\mathbf{M}}_{\mathrm{avg}},{\mathbf{M}}_{\max }]\in {\mathbb{R}}^{B\times 2\times H\times W}$ is obtained by concatenating the channel dimensions. Subsequently, a spatial weight map is generated through a single-layer 2D convolution:

$${\mathbf{a}}^{\left(s\right)}=\sigma \left({\mathrm{Conv}}_{k_s\times k_s}\left([{\mathbf{M}}_{\mathrm{avg}},{\mathbf{M}}_{\max }]\right)\right)\in {\mathbb{R}}^{B\times 1\times H\times W}$$

(5)

Here, $k_s$ represents the size of the spatial convolution kernel, which is used to cover a broader spatial context. Multiplying ${\mathbf{a}}^{\left(s\right)}$ element-wise with the original features yields the spatially re-calibrated features $\overline{\mathcal{X}}=\mathcal{X}\odot {\mathbf{a}}^{\left(s\right)}$. To achieve synergistic effects between channel-wise and spatial-wise attention, this paper employs parallel branching followed by gated linear fusion. Let $\lambda \in [0,1]$ denote the learnable fusion coefficient (which can be set as a scalar parameter shared across layers). The final output is then expressed as:

$$Y=\mathcal{X}\odot \left(\lambda ·\mathrm{Broadcast}\left({\mathbf{a}}^{\left(c\right)}\right)+(1-\lambda )·{\mathbf{a}}^{\left(s\right)}\right)+\mathcal{X}$$

(6)

The first term represents the reweighted additive residual after fusion, while the second term is the identity residual, which helps stabilize deep training and prevent excessive attention suppression. For finer-grained adaptation, $\lambda$ can also be dynamically generated via a lightweight $1\times 1$ convolution on the concatenation of $\mathrm{Broadcast}\left({\mathbf{a}}^{\left(c\right)}\right)$ and ${\mathbf{a}}^{\left(s\right)}$:

$$\lambda =\sigma \left({\mathrm{Conv}}_{1\times 1}\left([\mathrm{Broadcast}\left({\mathbf{a}}^{\left(c\right)}\right),{\mathbf{a}}^{\left(s\right)}]\right)\right)$$

(7)

At this point, the shape of $\lambda$ is consistent with ${\mathbf{a}}^{\left(s\right)}$, thereby achieving position-by-position gated fusion.

In summary, the dual-branch attention module establishes channel dependencies through ECA while highlighting key spatial regions via Spatial CBAM. Both components learn independently in parallel paths and collaborate during the fusion stage. This design achieves coordinated channel–spatial reweighting within a unified residual framework, providing more discriminative spatiotemporal feature representations for subsequent decoding and reconstruction.

2.2.2. DSC for Efficient Computation

To significantly enhance computational efficiency while preserving modeling capabilities, this paper introduces DSC [37] into the main convolutional block. DSC decomposes the traditional two-dimensional convolution operation (involving spatial feature extraction and channel fusion) into two steps: first performing spatial convolution independently on each input channel (Depthwise), then implementing a linear combination of the channel dimension using $1\times 1$ pointwise convolution (Pointwise). The computational process is as follows:

$$Y_{b,c_o,i,j}=\sum _{c_i=1}^{C_{\mathrm{in}}}\sum _{u=1}^K\sum _{v=1}^K{W}_{c_o,c_i,u,v}{X}_{b,c_i,i+u-\xi ,j+v-\xi }$$

(8)

where X and Y represent input and output features, respectively, W denotes the convolution kernel, and $(i,j)$, $(u,v)$ denote spatial coordinates. DSC first performs depthwise convolution:

$$Z_{b,c,i,j}=\sum _{u=1}^K\sum _{v=1}^K{W}_{c,u,v}^{\left(d\right)}{X}_{b,c,i+u-\xi ,j+v-\xi }$$

(9)

Each channel uses its own $K\times K$ kernel independently, with no inter-channel fusion. Pointwise convolution is then performed:

$$Y_{b,c_o,i,j}=\sum _{c=1}^{C_{\mathrm{in}}}{W}_{c_o,c}^{\left(p\right)}{Z}_{b,c,i,j}$$

(10)

Implement linear mixing of all channels using $1\times 1$ convolutions. Compared to the parameters of standard convolution $C_{\mathrm{in}}{C}_{\mathrm{out}}{K}^2$, DSC has parameters $C_{\mathrm{in}}{K}^2+C_{\mathrm{in}}{C}_{\mathrm{out}}$, significantly reducing model complexity. This makes it particularly suitable for lightweight design and small-sample inference scenarios.

This paper replaces all traditional $3\times 3$ convolutions in UNet with “Depthwise $3\times 3$ + Pointwise $1\times 1$” operations. The convolution block structure is “DSC + Batch Normalization (BN) + LeakyReLU,” followed by a tandem dual-branch attention mechanism. This design reduces parameter size and computational complexity while preserving the model’s ability to capture and distinguish the spatial structure of wind fields.

2.2.3. ASPP for Multi-Scale Context Aggregation

To enhance the model’s ability to perceive wind speed structures across different spatial scales, this paper introduces ASPP module into the bottleneck layer. Its workflow is illustrated in Figure 4.

ASPP achieves effective fusion of local and global context through multi-branch dilated convolutions (with varying dilation rates) and parallel feature aggregation via global average pooling. Each dilated convolution branch employs distinct dilation factors r, effectively expanding receptive fields without introducing additional parameter growth. The overall output undergoes fusion via $1\times 1$ convolutions:

$$F_{\mathrm{ASPP}}=\sigma \left({\mathrm{Conv}}_{1\times 1}\left(\mathrm{Concat}(F_1,\dots ,F_m,\mathrm{GAP})\right)\right)$$

(11)

Here, $F_{r_1}={\varphi }_{r_1}\left(\mathcal{X}\right)$ denotes the cavity convolutional branch with varying expansion rates, $F_{\mathrm{GAP}}$ represents the global pooling branch, Concat indicates channel concatenation, and $\sigma$ is the activation function. ASPP effectively enhances the model’s ability to extract complex multi-scale structures, providing rich contextual information for subsequent wind speed prediction.

2.2.4. Overall Architecture of DA-DSC-UNet

As illustrated in Figure 5, the DA-DSC-UNet framework is constructed upon the classic Encoder–Decoder architecture of UNet [38], specifically tailored to address the spatiotemporal heterogeneity inherent in coastal urban wind fields. The processing pipeline begins with an input tensor X, comprising a sequence of 12 historical wind velocity matrices stacked along the channel dimension.

In the encoding phase, the model employs a hierarchical structure with two levels of downsampling to progressively expand the receptive field while extracting high-level semantic abstractions. Within each feature extraction block, the proposed “DSC + Dual-Attention” mechanism is embedded. Specifically, the DSC layers efficiently extract local spatial patterns with minimal computational cost, while the subsequent Dual-Attention module adaptively recalibrates feature responses. This design enables the model to prioritize geographically significant structures—such as coastline inflections and high-density urban roughness elements—while suppressing irrelevant background noise.

At the bottleneck layer, where semantic information is most abstract, an ASPP module is deployed. By utilizing hybrid dilation rates ($r_1,r_2,r_3$) alongside a global average pooling branch, this module captures atmospheric processes across varying scales, ranging from local turbulence induced by urban morphology to regional sea-breeze circulations. The aggregated multi-scale features are then fused via a $1\times 1$ convolution to provide a comprehensive contextual representation.

Subsequently, the decoder reconstructs the high-resolution wind field through progressive upsampling. Crucially, skip connections are utilized to concatenate deep semantic features with the corresponding high-frequency details from the encoder, thereby preserving sharp gradients at the land–sea interface that might otherwise be lost during downsampling. These concatenated features undergo refinement through identical lightweight blocks to minimize artifacts. Finally, a projection layer maps the deep features to the output space, generating the predicted wind speed matrices for the next 12 time steps. Collectively, this architecture synergizes parameter efficiency with feature selectivity, offering a robust solution for real-time environmental monitoring in complex coastal zones.

2.3. Experimental Setup

This study employs 2021 observational data as the training set and data from January to February 2022 as the validation set. The selection of this timeframe was driven by two primary factors:

• Data Consistency: The multi-source satellite products (particularly FY-3E) provided the most stable and continuous high-quality observations during this period.
• Seasonal Representation: The training set covers a full annual cycle, enabling the model to learn seasonal wind patterns, while the validation set (Jan–Feb) represents the winter monsoon season, which is characterized by high wind speeds and volatility, serving as a rigorous test for model robustness.

The model input $X\in {\mathbb{R}}^{T\times C\times H\times W}$ is constructed by concatenating the normalized wind speed and wind direction matrices along the channel dimension, spanning 12 consecutive historical time steps. Although the primary forecasting target is the scalar wind speed (outputting a spatial wind speed matrix for the next 12 steps) due to its direct relevance to wind energy potential and disaster intensity, the inclusion of wind direction serves as an auxiliary kinematic feature. This allows the network to implicitly capture the vector evolution of air masses, thereby enhancing the accuracy of scalar speed prediction.

All models utilize Mean Squared Error (MSE) as the loss function, with the Adam optimizer selected and all parameters set to default values. The training epoch was set to 100, batch size to 512, and random seed fixed at 42 to ensure reproducibility. All experiments were conducted on Nvidia Tesla V100S GPUs. To comprehensively evaluate the proposed model’s performance, three experimental categories were established: First, comparative experiments pitted DA-DSC-UNet against several representative spatiotemporal prediction baselines: CNN-LSTM [39], ConvLSTM [23], Predictive Spatiotemporal Network (PSTN) [40], Dual Spatiotemporal Network (DSTN) [41], and Temporal Attention [42]. All comparison methods were trained and tested under identical data partitioning and hyperparameter tuning strategies to ensure fair comparison. To facilitate reproducibility and interpretability, the specific architectural configurations and hyperparameter settings for each model are detailed in

Table 3. Detailed hyperparameter configurations for the proposed model and baseline methods.

Model	Layers/Structure	Hidden Channels	Kernel Size	Learning Rate
CNN-LSTM	2 (CNN) + 2 (LSTM)	64, 128 (CNN); 128 (LSTM)	$3\times 3$	$1\times {10}^{-4}$
ConvLSTM	4 (Stacked Layers)	64, 64, 32, 32	$3\times 3$	$1\times {10}^{-3}$
PSTN	4 (ResNet Blocks)	32, 64, 128, 256	$3\times 3$	$5\times {10}^{-4}$
DSTN	3 (Encoder–Decoder)	64, 128, 64	$3\times 3$	$1\times {10}^{-3}$
Temporal Attention	2 (Self-Attention)	128 (Model Dim)	-	$1\times {10}^{-4}$
DA-DSC-UNet (Ours)	4 (Encoder) + Bridge	32, 64, 128, 256, 512	$3\times 3$ (DSC)	$1\times {10}^{-4}$

Note: All models were trained with a batch size of 512 and the Adam optimizer for 100 epochs.

. We ensured that the complexity of baseline models was adjusted to be commensurate with the input resolution and dataset scale. Second, ablation experiments were conducted by sequentially removing or replacing modules to quantify the contribution of each component. Third, robustness evaluation: Perturbed datasets were constructed to simulate real-world observational uncertainties and assess the model’s sensitivity to noise. This approach comprehensively demonstrates the model’s accuracy, stability, and generalization capabilities.

2.4. Evaluation Metrics

To comprehensively evaluate the multi-step wind speed prediction performance of the model, this paper selects Mean Absolute Error (MAE) and Root Mean Square Error (RMSE) as the main evaluation indicators. For each wind turbine location $(i,j)$ and the kth step prediction, the following definitions are used:

$$e_M{(i,j)}_k={\displaystyle \frac{1}{Q}}\sum _{q\in Q}\left|x_q{(i,j)}_k-{\widehat{x}}_q{(i,j)}_k\right|$$

(12)

$$e_R{(i,j)}_k=\sqrt{{\displaystyle \frac{1}{Q}}{\sum }_{q\in Q}{\left(x_q{(i,j)}_k-{\widehat{x}}_q{(i,j)}_k\right)}^2}$$

(13)

In this context, $x_q{(i,j)}_k$ and ${\widehat{x}}_q{(i,j)}_k$ represent the true value and predicted value of sample q at wind turbine location $(i,j)$ at step k, respectively, and Q denotes the total number of samples. To assess overall prediction performance, the average MAE (A-MAE) and average RMSE (A-RMSE) are further calculated for all wind turbines at each step:

$${\mathrm{A}\text{-}\mathrm{MAE}}_k={\displaystyle \frac{1}{m\times n}}\sum _{i=1}^m\sum _{j=1}^n{e}_M{(i,j)}_k$$

(14)

$${\mathrm{A}\text{-}\mathrm{RMSE}}_k={\displaystyle \frac{1}{m\times n}}\sum _{i=1}^m\sum _{j=1}^n{e}_R{(i,j)}_k$$

(15)

In addition, to evaluate the overall performance of the model in the multi-step prediction process, the average MAE and RMSE of all prediction steps are calculated, as shown in Equations (16) and (17). Here, h is the total number of prediction steps (12 in this study).

$$\overline{e_M}={\displaystyle \frac{1}{h}}\sum _{k=1}^h{\mathrm{A}\text{-}\mathrm{MAE}}_k$$

(16)

$$\overline{e_R}={\displaystyle \frac{1}{h}}\sum _{k=1}^h{\mathrm{A}\text{-}\mathrm{RMSE}}_k$$

(17)

3. Results

3.1. Overall Performance Comparison

To validate the model’s effectiveness in grid-based spatiotemporal wind field monitoring, we compared the proposed DA-DSC-UNet against five mainstream baseline models: CNN-LSTM, ConvLSTM, PSTN, DSTN, and Temporal Attention networks. The evaluation was conducted on a 12-step forecasting task, which is critical for short-term urban environmental monitoring. The quantitative results, measured by Average MAE (A-MAE) and Average RMSE (A-RMSE) for each prediction step k, are presented in

Table 4. Comparison of A-MAE (m/s) across different prediction models for 12 forecasting steps.

Step k	CNN-LSTM	ConvLSTM	PSTN 1	DSTN 2	Temporal Attention	DA-DSC-UNet
1	1.3350	1.3739	1.3804	1.4146	1.7512	1.1829
2	1.4823	1.5534	1.5349	1.5840	1.6317	1.2113
3	1.4654	1.5522	1.5473	1.5893	1.5666	1.1821
4	1.4481	1.5237	1.5298	1.5878	1.4779	1.1575
5	1.4387	1.5522	1.5213	1.6086	1.4292	1.1539
6	1.4302	1.5581	1.5307	1.5904	1.3663	1.1673
7	1.4195	1.5401	1.5316	1.5834	1.3074	1.1659
8	1.4288	1.5564	1.5138	1.5939	1.2563	1.1612
9	1.4518	1.5712	1.5432	1.6229	1.1972	1.1761
10	1.4432	1.5857	1.5363	1.6247	1.1623	1.1830
11	1.4437	1.5456	1.5299	1.5912	1.1329	1.1741
12	1.5161	1.5948	1.5742	1.6043	1.1032	1.1667
Average	1.4419	1.5423	1.5228	1.5829	1.3652	1.1735

¹ PSTN: Predictive Spatiotemporal Network. ² DSTN: Dual Spatiotemporal Network. Bold values indicate the best performance.

and

Table 5. Comparison of A-RMSE (m/s) across different prediction models for 12 forecasting steps.

Step k	CNN-LSTM	ConvLSTM	PSTN	DSTN	Temporal Attention	DA-DSC-UNet
1	1.8679	1.9028	1.9186	1.9655	2.4318	1.6908
2	2.0355	2.1230	2.0888	2.1652	2.2513	1.7254
3	2.0058	2.1111	2.0982	2.1715	2.1340	1.6646
4	1.9747	2.0988	2.0885	2.1867	2.0266	1.6396
5	1.9573	2.1321	2.0876	2.2143	1.9372	1.6600
6	1.9515	2.1185	2.0997	2.1893	1.8692	1.6599
7	1.9219	2.0988	2.0901	2.1699	1.7768	1.6535
8	1.9427	2.1384	2.0634	2.1842	1.7202	1.6491
9	1.9847	2.1607	2.1145	2.2248	1.6836	1.6730
10	1.9525	2.1557	2.0759	2.2054	1.6638	1.6966
11	1.9508	2.1026	2.0879	2.1795	1.6451	1.6671
12	2.0508	2.1656	2.1532	2.2022	1.6392	1.6703
Average	1.9663	2.1090	2.0805	2.1715	1.8982	1.6708

Bold values indicate the best performance.

, with visual comparisons illustrated in Figure 6.

As evidenced by the statistical metrics, DA-DSC-UNet demonstrates superior performance in reconstructing high-resolution wind fields. Specifically, the proposed model achieves the lowest error rates consistently across the first nine prediction steps, exhibiting exceptional capability in short-term forecasting. This advantage is particularly vital for coastal cities, where rapid weather changes require precise early warnings. Although the Temporal Attention model shows slightly lower errors in the extended forecast horizon (steps 10–12), due to its specialized temporal global attention mechanism, DA-DSC-UNet maintains the best overall performance. Aggregating across the entire forecast window, our model achieves an average MAE of 1.1735 m/s and an RMSE of 1.6708 m/s. This represents a substantial improvement over traditional spatiotemporal models; for instance, compared to DSTN and ConvLSTM, the MAE is reduced by 25.8% and 23.9%, respectively. Furthermore, the error curve of DA-DSC-UNet exhibits remarkable smoothness with minimal fluctuations, indicating effective suppression of error propagation during multi-step inference. This stability is attributed to the synergistic design of the DSC for efficient local feature extraction and the Dual-Attention mechanism for adaptive spatiotemporal recalibration. Consequently, the proposed framework offers a robust solution for real-time wind resource assessment and safety monitoring in complex urban–coastal environments.

3.2. Result Error Analysis

To further evaluate the operational reliability and generalization capability of the models under real-world conditions, we visualized the daily average error dynamics (absolute difference between ground truth and predicted values) using the validation dataset from January to February 2022. As illustrated in Figure 7, the DA-DSC-UNet model exhibits superior temporal consistency compared to the baseline models. While the proposed model may not strictly achieve the absolute minimum error at every single timestamp, its error curve consistently maintains a low-level trajectory with minimal variance throughout the observation period. In sharp contrast, comparison models such as CNN-LSTM, ConvLSTM, and PSTN display significant volatility. This instability is particularly pronounced during periods of intense wind speed fluctuations or extreme weather events, specifically on 5–6 January, 27–28 January, 13–14 February, and 25–27 February. During these intervals, the baseline models exhibit distinct error spikes, revealing their limitations in modeling complex, non-stationary spatiotemporal dynamics. Such susceptibility to abrupt atmospheric changes compromises their utility for reliable urban safety monitoring.

3.3. Spatiotemporal Error Distribution

To evaluate the potential of the proposed GeoAI model for urban environmental monitoring, Figure 8 visualizes the spatial distribution of MAE across the Fujian coastal zone. This region represents a typical complex environment where high-density urban agglomerations interact with the marine ecosystem, creating drastic changes in surface roughness and thermodynamic properties at the urban–ocean interface. As illustrated, DA-DSC-UNet exhibits superior spatial homogeneity, particularly in critical transitional zones where urban settlements and port infrastructure are concentrated. The model effectively suppresses error propagation in nearshore areas and estuarine cities (e.g., around the Min River Delta), maintaining a smooth error gradient despite the abrupt transition from open water to the built environment. This capability is crucial for assessing environmental risks, such as typhoon impacts on coastal urban infrastructure.

In contrast, baseline models like CNN-LSTM and ConvLSTM display widespread high-error bands along the coastline. These models struggle to model the nonlinear dynamics caused by the friction contrast between the sea surface and urban canopy, resulting in “blurring artifacts” in densely populated coastal strips. DSTN and PSTN, while improving locally, still retain fragmented error patterns in island–city transition zones. Consequently, DA-DSC-UNet demonstrates the most robust generalization capability, validating its utility as a precision tool for monitoring urban environmental dynamics and supporting sustainable coastal development.

3.4. Noise Robustness Evaluation

In operational urban environment monitoring, remote sensing data and ground observations are inevitably subject to signal degradation caused by complex electromagnetic interference, atmospheric scattering, or sensor calibration errors. To validate the reliability of the proposed GeoAI model under such non-ideal conditions, this study simulates varying degrees of observational uncertainty by introducing Gaussian noise to the input data. Specifically, zero-mean Gaussian noise $ϵ\sim \mathcal{N}(0,{\sigma }^2)$ was injected into the normalized input tensor $X_{norm}$ during the inference phase, as shown in Equation (18):

$$X_{noisy}=X_{norm}+ϵ,ϵ\in {\mathbb{R}}^{T\times C\times H\times W}$$

(18)

where the noise is applied to all feature channels C (affecting both wind speed and wind direction components) to simulate comprehensive sensor signal degradation. The noise standard deviation $\sigma$ ranges from 0 to 5.0, simulating scenarios extending from minor instrument noise to severe signal distortion during extreme weather events. Figure 9 illustrates the degradation trends of MAE and RMSE for all comparison models.

The results indicate that while predictive performance naturally declines with increasing noise intensity, the resilience of the models varies significantly. The proposed DA-DSC-UNet exhibits the most robust tolerance to data perturbations. Quantitatively, as $\sigma$ escalates from 0 to 5.0, the MAE of our model increases by approximately 1.25 m/s (RMSE by about 1.50 m/s). This suggests that the model’s internal attention mechanisms effectively filter out high-frequency noise, preserving the underlying physical features of the wind field.

In contrast, baseline models show higher sensitivity to input quality. For instance, the CNN-LSTM model exhibits a sharper deterioration, with its MAE increasing by approximately 1.70 m/s and its RMSE surging by over 2.25 m/s under extreme noise conditions. Notably, in scenarios with high interference ($\sigma =5.0$), the MAE of CNN-LSTM exceeds 3.1 m/s, whereas DA-DSC-UNet maintains it below 2.4 m/s. This instability in baseline models poses a risk for urban management, potentially leading to false alarms. The superior stability of DA-DSC-UNet ensures consistent and trustworthy monitoring support for urban infrastructure safety, even when input data quality is compromised by the complex urban environment.

3.5. Ablation Experiment

To quantitatively verify the contribution of each module to the spatiotemporal prediction task, systematic ablation experiments were conducted based on the UNet backbone. We investigated the specific impacts of ECA, CBAM, ASPP, and DSC on both predictive accuracy and computational efficiency.

Table 6. Quantitative comparison results of ablation experiments. Inference time and memory usage were tested on a single NVIDIA Tesla V100S with a batch size of 1 to simulate real-time inference scenarios. The best results are highlighted in bold.

Model	MAE	RMSE	Params	FLOPs	Infer. Time	Memory
	(m/s)	(m/s)		(M)	(ms)	(MB)
UNet (Baseline)	1.3512	1.7497	656,141	57.97	14.24	1024
ECA-UNet	1.3053	1.6966	656,153	58.04	14.85	1035
CBAM-UNet	1.3126	1.7022	656,533	58.04	15.10	1042
ECA+CBAM-UNet	1.2835	1.6901	656,545	58.11	15.42	1056
DSC-UNet	1.3501	1.7432	233,622	13.26	5.82	320
ASPP-UNet	1.2433	1.7162	1,132,045	61.35	22.56	1850
ECA+CBAM+ASPP-UNet	1.1857	1.6832	1,132,449	61.49	24.18	1920
ECA+CBAM+DSC-UNet	1.2776	1.6879	234,026	13.40	7.15	355
DA-DSC-UNet (Ours)	1.1735	1.6708	710,570	16.77	9.42	512

details the performance metrics, including MAE, RMSE, parameter quantity, and floating-point operations (FLOPs).

The results in Table 6 reveal the distinct role of each component in balancing accuracy and efficiency. To comprehensively assess deployment suitability, we additionally evaluated the Inference Time (latency per sample) and GPU Memory Usage during the inference phase:

• Impact of Attention Mechanisms: Both ECA and CBAM modules yield clear performance gains over the baseline. The ECA module reduces MAE by approximately 3.4%, while the dual-attention combination (ECA+CBAM-UNet) further suppresses the MAE to 1.2835 m/s. While attention mechanisms introduce a slight overhead in inference time (increasing from 14.24 ms to 15.42 ms), the cost is negligible compared to the accuracy gains.
• Efficiency of DSC: The DSC-UNet demonstrates a dramatic reduction in computational cost. By replacing standard convolutions, the parameter count drops by ∼64% (from 656 k to 233 k), and FLOPs decrease by ∼77% (from 57.97 M to 13.26 M), while maintaining accuracy comparable to the baseline. Crucially, the inference speed accelerates by nearly 2.5 times (5.82 ms vs. 14.24 ms), and memory usage drops to just 320 MB. This highlights the potential of DSC for lightweight deployment in resource-constrained urban monitoring devices.
• Importance of Multi-scale Context: The ASPP module (ASPP-UNet) significantly improves prediction accuracy (MAE drops to 1.2433 m/s) by capturing long-range dependencies. However, this comes at the cost of nearly doubling the parameter size and increasing computational load, with inference time rising to 22.56 ms.
• Synergy in the Proposed Model: The final DA-DSC-UNet achieves the optimal balance. By integrating the lightweight characteristics of DSC with the powerful feature extraction of dual attention and ASPP, our model achieves the lowest error rates (MAE = 1.1735 m/s, RMSE = 1.6708 m/s). Notably, compared to the standard UNet, our model improves MAE by 13.1% while consuming only 28.9% of the FLOPs (16.77 M vs. 57.97 M). Most importantly, despite the inclusion of the ASPP module, the efficient DSC backbone ensures that the inference time is maintained at only 9.42 ms, and memory usage is controlled at approximately 512 MB. This implies that the model can process over 100 frames per second on a standard GPU and fits comfortably within the memory limits of typical edge computing platforms (e.g., NVIDIA Jetson Nano), validating its suitability for practical real-time urban monitoring systems.

4. Discussion

4.1. Efficacy of Multi-Source Satellite Data Fusion

The core challenge in urban wind field monitoring lies in the mismatch between the spatial continuity required for analysis and the discrete nature of ground observations. Our results demonstrate that the proposed fusion strategy effectively bridges this gap. By integrating remote sensing data—which offers broad spatial coverage—with high-frequency ground station observations, the model overcomes the limitations of single-source methods.

Specifically, while ground stations provide high temporal resolution, they are often sparsely distributed and prone to “blind spots” in complex urban terrains. Satellite data complements this by providing a macroscopic view of regional wind patterns. The superior performance of DA-DSC-UNet, particularly in the ablation studies (Section 3.5), suggests that the model’s dual-attention mechanism successfully learns to weigh these heterogeneous inputs dynamically. ECA likely helps in selecting the most informative sensor features (e.g., prioritizing ground data when satellite signals are noisy), while CBAM focuses on localized wind anomalies caused by urban structures. This synergy ensures that the reconstructed wind field is not only accurate at observation points but also physically consistent across unmonitored areas.

4.2. Implications for Coastal Urban Management and SDG 11

The proposed GeoAI model holds significant practical value for the management of coastal cities, which are frequently exposed to complex wind environments and extreme weather events. Accurate wind field prediction is intrinsically linked to Sustainable Development Goal 11 (Sustainable Cities and Communities), particularly Target 11.5 (Disaster Risk Reduction) and Target 11.6 (Environmental Impact). By providing high-resolution spatiotemporal wind data, our framework directly supports these goals in three key aspects:

• Disaster Resilience and Safety: As demonstrated in the noise robustness evaluation (Section 3.4), our model maintains high stability even under significant signal interference. This reliability is critical for early warning systems during typhoons or convective storms, where sensor data is often degraded by heavy rainfall or transmission errors. Accurate wind field mapping can assist emergency responders in identifying high-risk zones for wind damage or falling debris.
• Real-time Environmental Monitoring: The lightweight nature of the DA-DSC-UNet (with only 16.77 M FLOPs and 0.71 M parameters) allows for direct deployment on resource-constrained edge computing devices. Unlike heavy numerical models that require supercomputing clusters, our framework can be integrated into portable meteorological drones or embedded edge chips (e.g., NVIDIA Jetson series). This enables on-board, real-time inference of satellite and sensor data with millisecond-level latency. This capability significantly reduces data transmission delays and supports real-time monitoring of pollutant dispersion, enabling rapid decision-making for urban traffic control or health advisories.
• Urban Planning and Energy: The high-resolution wind maps generated can inform the layout of high-rise buildings to mitigate the “urban canyon effect,” improving pedestrian comfort. Furthermore, the data can support the site selection for distributed small-scale wind energy harvesting systems in smart cities.

4.3. Limitations and Future Work

Despite the promising results, this study has certain limitations that outline directions for future research. First, the current model primarily focuses on 2D horizontal wind fields. However, the urban wind environment is inherently three-dimensional, with significant vertical wind shear influenced by building heights. The lack of vertical profile data limits the model’s ability to fully capture complex aerodynamic flows around skyscrapers. Future work will explore integrating LiDAR data or Computational Fluid Dynamics (CFD) simulations to extend the prediction to 3D space.

Second, while the model is robust to Gaussian noise, real-world sensor anomalies can be non-systematic and unpredictable (e.g., data gaps due to satellite revisit times). We plan to investigate Graph Neural Networks (GNNs) to better model the irregular topological relationships of sensor networks and handle missing data more effectively.

Finally, the current approach is purely data-driven, prioritizing computational efficiency and rapid inference for real-time urban monitoring. While this design significantly reduces the computational burden compared to traditional numerical models, it may occasionally generate predictions that lack strict adherence to physical laws (e.g., mass conservation). Future research will aim to bridge this gap by exploring hybrid paradigms. Specifically, we plan to incorporate lightweight physical constraints—inspired by PINNs—directly into the loss function rather than the network architecture. This strategy seeks to enhance the physical consistency of wind field reconstructions, particularly in sparse data scenarios, without sacrificing the model’s lightweight and deployment-friendly characteristics.

5. Conclusions

In the context of rapid global urbanization, accurate monitoring of environmental dynamics in coastal cities is pivotal for sustainable development and public safety. Addressing the challenge of resolving complex wind fields at the land–sea interface, this study proposed a novel GeoAI framework—DA-DSC-UNet—designed for high-resolution urban environment monitoring. By synergizing multi-source satellite data (ASCAT, FY-3E, QuickSCAT) with ground observations through dual attention mechanisms and multi-scale context modeling, the proposed model effectively captures the intricate spatiotemporal processes of urban wind fields that traditional methods often overlook. Comprehensive evaluations demonstrate that DA-DSC-UNet not only achieves state-of-the-art accuracy (MAE: 1.1735 m/s, RMSE: 1.6708 m/s) but also addresses the critical need for efficient processing in urban sensing networks. The integration of DSC reduces FLOPs by approximately 71% compared to standard architectures, making the model highly suitable for deployment on edge devices for real-time monitoring. Furthermore, its exceptional robustness under simulated noise confirms its reliability in processing data from diverse urban sensors, ensuring consistent performance even under non-ideal observational conditions. In summary, this research bridges the gap between massive multi-source remote sensing data and actionable urban environmental insights. By delivering precise, low-latency wind field information, this study contributes significantly to the realization of Sustainable Development Goal (SDG) 11, offering valuable scientific support for data-driven urban management and disaster mitigation strategies in complex coastal zones.