This study focuses on the Gaoqiao Mangrove National Nature Reserve, located on the northeastern coast of the Beibu Gulf in Guangdong Province, China (21°35’N, 110°20’E), encompassing Gaoqiao Town, Lianjiang City, and adjacent buffer zones (Figure 1). As the northernmost mangrove distribution in mainland China, the reserve is a globally significant site for mid-latitude mangrove ecosystems (Kathiresan and Bingham, 2001). It is the largest contiguous mangrove protected area along China’s continental coastline, supporting exceptional biodiversity, including 26 native species (e.g., Bruguiera gymnorrhiza, Kandelia obovata, Rhizophora stylosa), five introduced species (e.g., Sonneratia apetala), and diverse aquatic taxa (comprising 97 phytoplankton and 159 benthic diatom species) (Jia et al., 2018; Liu et al., 2024). Beyond its ecological significance, the Gaoqiao mangrove ecosystem provides substantial economic value, contributing to local livelihoods through fisheries, aquaculture, and ecotourism, with mangroves globally estimated to support economic benefits ranging from $2,000 to $9,000 per hectare annually (Salem and Mercer, 2012; Malik et al., 2015). The region exhibits a southern subtropical maritime monsoon climate, with an average annual temperature of 23.1°C, precipitation of 1600–1900 mm, 2100 sunshine hours, and an average of 2.3 typhoons annually during summer and autumn (Jia et al., 2018).

This study utilized multi-temporal Landsat surface reflectance products (Level-2) acquired from the United States Geological Survey (USGS) EarthExplorer platform to examine the spatiotemporal dynamics of the Gaoqiao Mangrove National Nature Reserve (21°33’N, 109°44’E), located at the Guangdong–Guangxi border, China, from 1993 to 2023. The dataset consists of Landsat 5 Thematic Mapper (TM, 1993–2012) and Landsat 8 Operational Land Imager (OLI, 2013–2023) images, providing a consistent 30 m spatial resolution. Only the three visible bands (blue, green, red) were used as input to the deep learning models.

A total of 127 scenes were initially available for the study period. After excluding 16 scenes with excessive cloud cover or sensor artifacts, 111 high-quality scenes were retained to ensure complete annual coverage across the 31-year record (Table 1).

For independent validation of the 2023 mangrove extraction, high-resolution drone imagery (3 cm/pixel, RGB + NIR bands) was acquired on 7 November 2023 using a DJI Matrice M300 platform, covering 4.35 km² of the reserve (Figure 2). This drone survey provided ground-truth reference for confusion matrix analysis (Table 2).

All Landsat scenes were clipped to the reserve boundary plus a 2 km buffer and co-registered to a common UTM Zone 49N grid (RMSE < 0.5 pixel) using ENVI 5.6 and Python scripts.

The workflow adopts a two-stage design:

Stage 1 – Mangrove extraction (spatial segmentation task)

From the 111 preprocessed annual scenes spanning the entire study period (1993–2023), 1280 overlapping patches (256 × 256 pixels, 50% overlap) with corresponding manually delineated binary mangrove masks were generated. This multi-temporal pooling strategy ensures the model captures diverse phenological and tidal conditions. Extensive online data augmentation (random rotation, horizontal/vertical flipping, brightness ±25%, contrast adjustment) was applied during training. The 1280 patches were randomly shuffled and split in a 7:2:1 ratio:

Random splitting is appropriate for this stage because the segmentation model processes each patch independently without temporal context.

Stage 2 – Spatiotemporal forecasting (temporal prediction task)

The trained U-Net+SE+CBAM model was applied to all 111 full scenes (1993–2023), producing one binary mangrove mask per year (31 annual masks total). These masks were chronologically divided as follows:

This strict temporal split prevents future information leakage. Forecasting models use sliding windows of five consecutive annual masks to predict the subsequent year’s mask, with three-year forecasts (2024–2026) generated recursively from the 2019–2023 window.

The extraction of mangrove distributions in this study relies on an enhanced U-Net model integrated with Squeeze-and-Excitation (SE) and Convolutional Block Attention Module (CBAM) to process Landsat imagery (1993–2023) from Landsat 5 TM and Landsat 8 OLI. The U-Net framework employs an encoder-decoder architecture for semantic segmentation, addressing the challenge of low spatial resolution (30 m) that often blurs mangrove boundaries and causes misclassification in complex coastal environments. The encoder downsamples the input through convolutional and pooling layers to extract multi-scale features, such as mangrove patches and tidal zones, while the decoder, supported by skip connections, upsamples these features to restore spatial detail, ensuring precise delineation despite resolution limitations.

To enhance feature discrimination, the SE module targets spectral noise interference, a common issue in coastal regions where reflections from water bodies or sediments obscure vegetation signals. SE applies global average pooling to compress spatial dimensions into a channel descriptor, followed by a fully connected layer and sigmoid activation to recalibrate feature responses, prioritizing informative bands (e.g., red and near-infrared) for mangrove health while suppressing non-vegetation noise. This refinement boosts segmentation accuracy, contributing to a notable performance uplift.

Complementing SE, the CBAM module improves edge detection accuracy, tackling challenges from tidal influences and adjacency to water bodies. CBAM uses a two-stage approach: a channel attention module weights feature importance to enhance mangrove signals and suppress tidal noise, while a spatial attention module highlights pixel-level differences to refine mangrove edges, reducing misclassification in dynamic intertidal zones. This dual mechanism enhances the model’s robustness across varying conditions.

Traditional approaches to mangrove area forecasting have relied on statistical models (e.g., Markov chains, ARIMA) or simple regression of historically mapped areas, which typically ignore spatial patterns and biological constraints, leading to unrealistic long-term projections in dynamic coastal systems. This study addresses these limitations through a two-stage deep learning pipeline: high-quality annual binary masks (1993–2023) are first generated by the U-Net+SE+CBAM segmentation model (Section 2.3), and these masks are then used as input for sequence-based forecasting. The primary recommended architecture is U-Net–ConvLSTM, which effectively combines U-Net’s precise spatial feature extraction with ConvLSTM’s efficient temporal modelling, offering an excellent balance of accuracy, simplicity, and computational efficiency. To explore whether additional ecological regularization can further constrain unrealistic long-term trends, an optional asymmetric Ecological Constraint Loss (ECOLOSS) is introduced, yielding the ConvLSTM+ECOLOSS variant that penalizes annual growth rates outside the regionally observed range of approximately –5% to +10% (Section 2.4.2). Performance of U-Net–ConvLSTM and ConvLSTM+ECOLOSS, along with five additional baselines (standalone ConvLSTM, ConvLSTM+Self-Attention, 3DCNN, LSTM, and RNN), is evaluated below.

Both the mangrove extraction and spatiotemporal forecasting models were trained on an NVIDIA 4090D GPU with 16 GB memory.

For mangrove extraction (U-Net+SE+CBAM), 1280 overlapping patches (256 × 256 pixels) were generated from the 111 annual scenes. Extensive online data augmentation (random rotation, horizontal/vertical flips, brightness/contrast adjustment) was applied. The patches were randomly split in a 7:2:1 ratio (70% training, 20% validation, 10% test). Training used binary cross-entropy loss and the Adam optimizer (initial learning rate 0.001) for a maximum of 100 epochs with early stopping (patience = 10). An ablation study for the attention modules is presented in Table 4.

For spatiotemporal forecasting, the 31 annual binary masks were chronologically divided into training (1993–2015), validation (2016–2018), and test (2019–2023) periods. All sequence models used a 5-year sliding window of masks ([5, 256, 256, 1]) to predict the subsequent year and were trained with Dice loss and the Adam optimizer (initial learning rate of 10−4, batch size 2) for a maximum of 100 epochs with early stopping (patience = 10). The ConvLSTM+ECOLOSS variant additionally incorporated the asymmetric ecological constraint loss (Section 2.4.2, λ = 0.5). No transfer learning was employed due to the consistent resolution of Landsat imagery throughout the study period.

Forecasting performance was evaluated using Intersection over Union (IoU) and F1-score for spatial accuracy, and Mean Absolute Error (MAE) and Root Mean Square Error (RMSE) for predicted mangrove area. Ablation analysis for ECOLOSS is presented in Table 3.

All experiments were implemented in Python using the PyTorch framework on an NVIDIA 4090D GPU with 16 GB memory. The code is available from the corresponding author upon reasonable request. Prediction results were visualized as line graphs of mangrove area (hectares) over time, with the U-Net+SE+CBAM mapping results serving as the reference trajectory.

To evaluate mangrove extraction performance in the Gaoqiao Mangrove National Nature Reserve, the proposed U-Net+SE+CBAM model was compared with traditional NDVI thresholding (threshold = 0.3) and ablation variants. Reference masks were derived from high-resolution Google Earth historical imagery (1993–2023) and the Global Mangrove Watch dataset.

The U-Net+SE+CBAM model integrates Squeeze-and-Excitation (SE) modules for channel-wise feature recalibration and Convolutional Block Attention Module (CBAM) for spatial attention, enabling effective differentiation of mangroves from tidal flats and water bodies in complex coastal environments (IoU = 0.815 and F1-score = 0.928, see Table 5; overall accuracy = 98.55% as confirmed by the confusion matrix in Table 2). In contrast, NDVI thresholding is more sensitive to residual cloud cover and tidal variation, while U-Net+SE improves spectral feature mapping but shows weaker boundary delineation in fragmented patches. The Gaoqiao region, the northernmost native mangrove ecosystem in mainland China and home to 26 true mangrove species, particularly benefits from such high-precision mapping to support evidence-based conservation.

An ablation study was conducted to evaluate the contributions of the SE and CBAM modules (Table 4). The baseline U-Net achieved an IoU of 0.648 and an F1-score of 0.779. Adding SE alone improved performance to IoU = 0.663 (relative improvement of 2.4% over baseline) and F1-score = 0.791, while CBAM alone yielded IoU = 0.629 and F1-score = 0.764. The full model combining both attention mechanisms (U-Net+SE+CBAM) achieved the highest performance with IoU = 0.718 (relative improvement of approximately 10.6% over the baseline U-Net and 8.3% over U-Net+SE alone) and F1-score = 0.822. This synergistic effect demonstrates the complementary roles of channel-wise recalibration (SE) and spatial attention (CBAM) in reducing misclassification, particularly in tidally influenced zones with mixed water-vegetation signals.

These quantitative findings are further detailed in the confusion matrix derived from independent 2023 reference data (Table 2). The U-Net+SE+CBAM model exhibits superior performance with an overall accuracy (OA) of 98.55% and a Kappa coefficient of 0.9708, while NDVI thresholding achieves OA = 91.92% and Kappa = 0.8388. The high producer’s accuracy (PA = 98.78%) and user’s accuracy (UA = 98.48%) for mangroves in U-Net+SE+CBAM underscore its precision in detecting mangrove areas, compared to NDVI’s PA = 87.54% and UA = 96.97%. This validates the model’s reliability for detailed mangrove mapping in the Gaoqiao region. Qualitative insights into these methods are presented in Figure 4.

The spatiotemporal forecasting performance was evaluated on the 31 annual binary mangrove masks (1993–2023) generated by the U-Net+SE+CBAM model (Section 2.3). Historical trends derived from these masks show a decline from approximately 700 ha in 1993 to a minimum of 650 ha around 2008, followed by gradual recovery to 675 ha by 2023, consistent with documented phases of aquaculture expansion and subsequent conservation efforts.

As described in Section 2.3, the masks were chronologically split into training (1993–2015), validation (2016–2018), and test (2019–2023) periods. All sequence models used a 5-year rolling window to predict the subsequent year. This window length was selected as it yielded the lowest validation error among tested lengths of 3–7 years.

A qualitative comparison for September 2019 is shown in Figure 5. The reference mask from U-Net+SE+CBAM (a) accurately resolves narrow tidal-creek fringes and fragmented patches. LSTM (b) severely over-generalizes, ConvLSTM (c) sharpens boundaries but introduces occasional artificial expansion into mudflats, and the U-Net–ConvLSTM baseline (d) closely matches the reference in core areas while merging some thin features. The ConvLSTM+ECOLOSS prediction is visually very similar to the standalone ConvLSTM result (c) for this single year. This similarity is expected, as the primary role of the asymmetric ECOLOSS is to suppress biologically implausible multi-year runaway growth or decline, rather than to modify short-term spatial patterns that already lie within realistic bounds.

Temporal trajectories over the test period are presented in Figure 6. The reference series exhibits mild recovery from 672 ha in 2019 to 675 ha in 2023. U-Net–ConvLSTM and ConvLSTM+ECOLOSS track this trend most faithfully, with only minor deviations. Other convolutional models show slight over-prediction in 2021–2022, while purely temporal models deviate substantially.

Quantitative metrics on the test period (2019–2023) are summarized in Table 6. The primary recommended U-Net–ConvLSTM model achieves strong performance (IoU = 0.787, F1-score = 0.874, MAE = 6.83%, RMSE = 9.03%). Adding the optional asymmetric ECOLOSS regularization yields ConvLSTM+ECOLOSS with the highest numerical scores (IoU = 0.793, MAE = 6.70%), but the absolute improvement remains marginal — an increase of 0.006 in IoU and a reduction of 0.13 percentage points in MAE — and is consistent across all metrics. Ablation analysis (Table 3) confirms that ECOLOSS primarily serves as an ecological safeguard, providing a modest reduction in extreme annual growth-rate outliers relative to unconstrained or symmetric variants, rather than acting as a major driver of short-term accuracy.

To generate near-term forecasts of mangrove dynamics in the Gaoqiao region (2024–2026), three models were employed: the primary recommended U-Net–ConvLSTM model, the pure ConvLSTM model, and the ConvLSTM+ECOLOSS variant incorporating asymmetric ecological regularization. All models were driven recursively using a 5-year sliding window on the annual mangrove masks derived from the U-Net+SE+CBAM segmentation for the historical period 1993–2023. Over this historical span, mangrove area increased from 542.25 ha in 1993 to 693.45 ha in 2023, representing a net gain of 151.20 ha (+27.88.

The predicted area trajectories indicate continued gradual recovery over the three-year forecast horizon, although the models differ in projected magnitude and rate of growth. The U-Net–ConvLSTM model forecasts mangrove areas of 700.38 ha in 2024, 708.30 ha in 2025, and 710.46 ha in 2026, yielding a modest cumulative increase of 10.08 ha (+1.44% relative to the 2024 baseline) and the most conservative trajectory overall. The ConvLSTM+ECOLOSS variant projects slightly higher values of 682.83 ha, 697.77 ha, and 701.37 ha for the corresponding years, resulting in a cumulative gain of 18.54 ha (+2.72%). The pure ConvLSTM model produces the strongest expansion, with predicted areas of 681.21 ha, 702.63 ha, and 710.19 ha, corresponding to a cumulative increase of 28.98 ha (+4.25%).

The full historical time series (1993–2023) and the near-term forecasted trajectories (2024–2026) for the three models are presented in Figure 7, with the left panel showing the complete series from 1993 to 2026 and the right panel providing a zoomed-in view of the prediction period.

Ablation analysis demonstrates that the pure ConvLSTM exhibits a pronounced tendency toward rapid short-term expansion, whereas incorporation of the asymmetric ECOLOSS provides mild moderation of this trend (reducing the three-year increase from +4.25% to +2.72%), confirming a constrained but limited regulatory effect that does not markedly reshape the overall forecast amplitude. By comparison, the U-Net–ConvLSTM model delivers the most stable and ecologically plausible predictions, owing to its integrated architecture that effectively couples detailed spatial feature extraction with robust temporal modeling. Its conservative growth trajectory aligns closely with known biophysical limitations in the Gaoqiao Reserve, including restricted seaward extension due to tidal exposure and primary expansion potential confined to northern intertidal zones.

These near-term forecasts suggest that, under continuing current management practices, mangrove recovery in the Gaoqiao Reserve will remain gradual and biologically reasonable through 2026, with total extent stabilizing near 700 ha. The U-Net–ConvLSTM model is considered the most reliable benchmark for these projections due to its superior stability and fidelity to observed ecological constraints.