High-Resolution Mapping Coastal Wetland Vegetation Using Frequency-Augmented Deep Learning Method

Highlights

• First use of very-high resolution (2 cm) image data for Mapping Coastal Wetland Vegetation: We have used drone imagery and deep learning techniques to complete the task of fine tuned classification of Wetland Vegetation.
• A method for coastal vegetation classification from high resolution imagery is proposed: In this paper, we proposal a augment frequency-domain features network—AFDFNet. Experimental results demonstrate that AFDFNet consistently outperforms existing deep learning models, achieving state-of-the-art performance.
• An Multi-scale Feature Enhancement Module: We designed a multi-scale feature enhancement module to compensate for the misclassification phenomenon caused by the lack of frequency domain features and contextual information in the network.

Abstract

Coastal wetland vegetation exhibits pronounced spectral mixing, complex mosaic spatial patterns, and small target sizes, posing considerable challenges for fine-grained classification in high-resolution UAV imagery. At present, remote sensing classification of ground objects based on deep learning mainly relies on spectral and structural features, while the frequency domain features of ground objects are not fully considered. To address these issues, this study proposes a vegetation classification model that integrates spatial-domain and frequency-domain features. The model enhances global contextual modeling through a large-kernel convolution branch, while a frequency-domain interaction branch separates and fuses low-frequency structural information with high-frequency details. In addition, a shallow auxiliary supervision module is introduced to improve local detail learning and stabilize training. With a compact parameter scale suitable for real-world deployment, the proposed framework effectively adapts to high-resolution remote sensing scenarios. Experiments on typical coastal wetland vegetation including Reeds, Spartina alterniflora, and Suaeda salsa demonstrate that the proposed method consistently outperforms representative segmentation models such as UNet, DeepLabV3, TransUNet, SegFormer, D-LinkNet, and MCCA across multiple metrics including Accuracy, Recall, F1 Score, and mIoU. Overall, the results show that the proposed model effectively addresses the challenges of subtle spectral differences, pervasive species mixture, and intricate structural details, offering a robust and efficient solution for UAV-based wetland vegetation mapping and ecological monitoring.

1. Introduction

Coastal wetlands are among the most ecologically valuable ecosystems on Earth. They play indispensable roles in maintaining biodiversity, sequestering carbon, mitigating emissions, preventing shoreline erosion, and regulating regional climate systems [1]. As a key component of these ecosystems, wetland vegetation not only reflects their structural and functional attributes but also provides critical indicators for ecological succession, invasive species expansion, and restoration processes [2].

Traditional investigations of coastal wetland vegetation have largely relied on field surveys, which are expensive, spatially and temporally constrained, and relatively inefficient [3,4,5]. Remote sensing offers an effective alternative, enabling large-scale and long-term monitoring while reducing labor and resource requirements. Among various remote sensing platforms, unmanned aerial vehicles (UAVs) have gained widespread attention due to their high mobility, low cost, and exceptional spatiotemporal resolution. UAV-based remote sensing has therefore been increasingly applied in ecological monitoring, agricultural management, and wetland assessment [6]. Previous studies have demonstrated the potential of UAV imagery in coastal vegetation research, including:

Akinaga et al. used high-resolution drone imagery to assess blue carbon reserves in tidal flat seagrass beds [7]. Liu et al. applied deep learning methods to the classification of coastal habitats based on drones and compared different data fusion and semantic segmentation techniques [8]. James et al. explored the modeling and prediction of coastal vegetation and environmental information from near-infrared (NIR) UAV imagery, which has reference value for vegetation indices and classification [9]. Morgan et al. studied how to estimate the aboveground biomass (AGB) of intertidal wetland vegetation using UAV multispectral imagery combined with ground samples [10]. While UAV remote sensing provides abundant high-resolution data, the sheer volume of such imagery poses considerable challenges for data processing and analysis. To process massive amounts of high-resolution image data more efficiently, computer vision techniques are becoming increasingly dominant.

Advancements in computer vision have greatly propelled remote sensing image analysis. From early digital image processing techniques to modern deep learning-based approaches, these methods have proven highly effective for UAV imagery. Deep learning, in particular, eliminates the need for handcrafted feature design and offers strong representational capacity, making it the dominant paradigm in current remote sensing research. Compared with traditional digital image processing methods, deep learning-based approaches exhibit distinct advantages in coastal wetland studies. Conventional methods often rely on manually designed features and fixed thresholds, which are sensitive to illumination variations, seasonal changes, and the complex spectral–spatial heterogeneity commonly observed in coastal wetland environments. In contrast, deep learning models are capable of automatically learning hierarchical and discriminative features from data, enabling more robust representation of diverse wetland components such as vegetation, tidal flats, and water bodies. Architectures such as convolutional neural networks (CNNs), attention mechanisms, and vision transformers have shown remarkable performance in coastal vegetation classification and recognition [11,12]. Wang et al. achieved high-precision pixel-level mangrove species classification using a semantic segmentation network [13]. Ke et al. conducted detailed mapping and change detection in the Liaohe Estuary using time-series imagery and deep learning models [14]. Cruz et al. identified an invasive species English glasswort in salt marsh and mudflat environments using a U-Net architecture built on an Inception-v3 backbone [15]. These studies collectively highlight the effectiveness of deep learning for species extraction and recognition in coastal wetlands. However, most existing methods focus primarily on spatial textures and spectral features, while largely overlooking the rich information embedded in images’ frequency domain.

Although numerous studies have advanced coastal wetland monitoring at regional and temporal scales, most large-scale or long-term analyses still rely on moderate- and low-resolution satellite imagery such as Landsat, MODIS, and Sentinel-2 [16,17,18,19,20]. These datasets provide long-term records, global coverage, and free accessibility, making them well suited for time-series analysis and large-scale mapping [21]. As a result, many national, regional, and global wetland mapping products rely on imagery with spatial resolutions of 10–30 m. In contrast, studies utilizing UAV or commercial sub-5 m and sub-meter resolution imagery are typically localized case studies used to validate, refine, or interpret heterogeneity in medium-resolution products and have not yet become mainstream [22].

In summary, existing research on coastal wetland vegetation classification is still dominated by medium and low resolution satellite imagery. Studies leveraging ultra-high-resolution UAV imagery remain limited, and most current methods overlook critical frequency-domain characteristics. Consequently, these approaches often struggle with small vegetation targets and complex co-occurrence patterns. To address these challenges, this study constructs an ultra high-resolution RGB UAV dataset for coastal vegetation classification, including representative species such as Reeds, Spartina alterniflora, Suaeda salsa, and water bodies. Furthermore, we proposal a augment frequency-domain features network—AFDFNet. Experimental results demonstrate that AFDFNet consisently outperforms existing deep learning models, achieving state-of-the-art performance.

2. Materials and Methods

2.1. Dataset Construction

The data used in this study were acquired using an airborne UAV manufactured by Feima Robotics (Beijing, China). The UAV captured true orthophoto imagery with a spatial resolution of 0.02 m using an RGB camera. After preprocessing steps including stitching and mosaicking, the resulting image measured 22,169 × 20,732 pixels and was projected onto the WGS-84 coordinate system. The processed imagery and the geographic extent of the study area are shown in Figure 1.

This drone survey was conducted in the coastal wetlands of Dongying City, Shandong Province, China, covering an area of approximately 200 square meters. Dongying is located at the mouth of the Yellow River and is a typical delta wetland. Vegetation in this region exhibits distinct zonation patterns. Suaeda salsa primarily grows in low-lying saline–alkaline zones of the intertidal area and is one of the most representative halophytic communities in the Yellow River Delta [23]. Reeds is widely distributed across freshwater and slightly brackish wetlands and serves as the dominant native species in the region [24]. The invasive species Spartina alterniflora has rapidly expanded along the coastal zone, forming complex competitive and coexisting relationships with Reeds in certain areas.

Figure 2 shows the constructed dataset. The selected images fully cover all major vegetation types within the study area. Pixel-level labels were manually annotated through visual interpretation. The dataset covering six categories: Reeds, Spartina alterniflora, Suaeda salsa, weeds, water bodies, and bare tidal flats. Suaeda salsa, as a beneficial vegetation for wetland environmental management, exhibited weakness, resulting in the largest sample size. Next, Phragmites australis and Spartina alterniflora represented native and invasive species in the study area, reflecting vegetation invasion in wetlands, hence their smaller sample sizes. Other less distributed vegetation was grouped as weeds, with water bodies and bare beaches serving as the background; these species had relatively smaller sample sizes. Specifically, A total of 160,256 × 256 pixel image-label pairs were generated, covering six major categories: Reeds, Spartina alterniflora, Suaeda salsa, weeds, water bodies, and bare tidal flats. Spartina alterniflora, Suaeda salsa, weeds, are three types of vegetation that are important for the wetland ecosystem. Therefore, the dataset was divided into 32 images each represent reeds and Spartina alterniflora; 48 images represent Suaeda salsa, which is a weak target in this area; 16 images represent weeds; and 16 images represent water bodies and exposed tidal flats, which are mostly background. The dataset was divided into training and testing sets in a 7:3 ratio.

2.2. AFDFNet Model Development

2.2.1. Overall Framework

As a typical transitional ecosystem, coastal wetlands exhibit highly mixed spectral characteristics. Vegetation types such as Reeds, Spartina alterniflora, and halophytic plants often coexist and intermingle within the same area. Their canopy structures and moisture conditions also change dynamically. In addition, Suaeda salsa pixels frequently contain mixed spectral responses from adjacent tidal flats, making it difficult for traditional remote sensing imagery to maintain spectral purity at the pixel level. Based on the dataset constructed in Section 2.1, a representative image is selected for each category, and its mean at the pixel scale is calculated to obtain the spectral curve. The spectral curves obtained from UAV imagery are shown in Figure 3.

Moreover, the characteristic coexistence and mosaic patterns of coastal wetland vegetation further blur class boundaries. For instance, Reeds and Spartina alterniflora often appear in mixed stands along the wetland ecotone, leading to overlapping textures and spectral signatures among vegetation types. Such interactions reduce inter-class separability and pose significant challenges for conventional classification methods [25,26]. To address these issues, we propose AFDFNet, a deep learning framework specifically designed for UAV-based coastal wetland vegetation classification.

The proposed method is illustrated in Figure 4. The architecture adopts an encoder–decoder structure, with ResNet-50 serving as the backbone network [27]. ResNet-50, which achieved first place in the ImageNet 2015 classification task, remains one of the most widely used feature extractors for high-resolution remote sensing imagery. The backbone extracts four hierarchical feature maps from shallow to deep levels. These feature maps are then fed into the Multi-scale Feature Enhancement Module (MFEM), which contains two parallel branches: A large-kernel convolution branch, designed to capture contextual dependencies among different vegetation types. A frequency-domain interaction branch, used to extract informative frequency components that are typically overlooked in conventional spatial-domain models. After processing through the MFEM, the enhanced features are progressively upsampled from deep to shallow layers and fused via skip connections. The final output is upsampled to the original image resolution for loss computation.

2.2.2. Multi-Scale Feature Enhancement Module (MFEM)

In coastal wetland vegetation classification, incorporating contextual modeling and frequency-domain feature extraction is essential. On the one hand, many vegetation patches in wetlands (e.g., Suaeda salsa) exist as small or weakly expressed targets [28], making it difficult for conventional local convolutions to achieve adequate discriminative capability. On the other hand, mixed stands and mosaic patterns are widespread in coastal wetlands, and UAV imagery typically provides limited spectral dimensionality. As a result, single-pixel spectral and textural information is often insufficient. In this context, frequency-domain features can effectively separate low-frequency regional structures from high-frequency boundary details [29], thereby improving the model’s sensitivity to ambiguous boundaries and mixed pixels.

To address the limitations of traditional wetland vegetation classification methods—which often overlook frequency-domain and contextual information—we design an embedded Multi-scale Feature Enhancement Module (MFEM), shown in Figure 5. The MFEM consists of two parallel branches: a large-kernel convolution branch and a frequency-domain interaction branch. This design reduces misclassification and omission errors caused by insufficient contextual or frequency cues in standard deep networks.

(1)

Large-kernel Convolution Branch

To enhance contextual modeling capability, this branch employs large convolution kernels to expand the receptive field and capture richer spatial relationships. Specifically, the feature maps f extracted from the backbone are processed through two large-kernel convolutions (5 × 5 and 7 × 7) to extract multi-scale spatial features. The resulting feature maps are concatenated along the channel dimension and passed through both max pooling and average pooling to obtain statistical descriptors at two scales. These pooled features are then fed into a 7 × 7 convolution to generate two spatial attention masks, which modulate the outputs of the 5 × 5 and 7 × 7 convolutions through element-wise multiplication. Finally, the enhanced features are fused and forwarded to the next stage.

To preserve the fine spatial details inherent in ultra-high-resolution UAV imagery, the feature map resolution remains unchanged throughout this branch.

The process can be mathematically expressed as:

$$\begin{array}{c}L={conv}_{5 \times 5}\left(f\right)\times m_1+{conv}_{7 \times 7}\left(f\right)\times m_2\end{array}$$

(1)

where

• $L$ denotes the output of the large-kernel branch,
• $f$ is the backbone feature map,
• $m_1$ and $m_2$ represent spatial attention masks generated via average pooling and max pooling, respectively.

(2)

Frequency-domain Interaction Branch

This branch aims to strengthen representation learning from a frequency-domain perspective. Starting from the same backbone feature map f, the branch performs explicit separation of high-frequency and low-frequency components to enable multi-frequency feature modeling.

A 1 × 1 convolution is first applied to extract high-frequency information while preserving spatial resolution. Meanwhile, low-frequency features are obtained by applying average pooling followed by a 1 × 1 convolution to the downsampled feature map. Based on these decomposed components, an Octave Convolution [30] is used to enable bidirectional interaction and fusion between high- and low-frequency signals. Octave Convolution (OctConv) is designed to reduce spatial redundancy in convolutional neural networks by explicitly decomposing feature maps into high- and low-frequency components. Instead of processing all feature channels at the same spatial resolution, OctConv allocates high-frequency features to full-resolution maps to preserve fine spatial details, while low-frequency features are computed at a reduced resolution to capture global contextual information. Information exchange between the two frequency components is achieved through four convolutional pathways, including high-to-high, high-to-low, low-to-low, and low-to-high transformations, where downsampling and upsampling operations are used to enable cross-frequency interaction. By performing convolution on low-frequency features at a coarser spatial scale, OctConv effectively enlarges the receptive field and reduces computational cost, while maintaining the ability to model both detailed structures and large-scale semantic patterns. As a learnable operator, the Octave Convolution adaptively adjusts cross-frequency information flow during training, achieving effective fusion of complementary spectral–spatial cues.

The computation is formulated as:

$$\begin{array}{c}F={OctaveConv(conv}_{1 \times 1}\left(f\right)+{conv}_{1\times 1}\left[downsample\right(f\left)\right])\end{array}$$

(2)

where

• $F$ is the output of the frequency-domain interaction branch.
• $f$ is the backbone feature map.

(3)

Decoder with Cascaded MFEMs

Multiple MFEMs are stacked to form the decoder of the proposed model. The four feature maps enhanced by MFEM are progressively upsampled from deep to shallow layers. At each stage, the upsampled feature map is passed through a 3 × 3 convolution and concatenated with the corresponding MFEM output from the next layer. This process repeats until the feature map resolution matches that of the input image.

The decoding process can be expressed as:

$$\begin{array}{c}{D}_{i+1}={(L}_i+F_i)+D_i\end{array}$$

(3)

$$\begin{array}{c}{D}_0={(L}_0+F_0)\end{array}$$

(4)

where

• $D_i$ denotes the decoder features at each scale.

2.2.3. Loss Function

To enhance the model’s ability to recognize fine-grained structures of coastal vegetation, this study adopts a joint loss function consisting of two components. The first component is the conventional semantic segmentation loss, where pixel-wise cross-entropy is applied to constrain the final semantic prediction. This loss guides the model to accurately learn the spatial distribution patterns of different vegetation types and serves as the primary supervisory signal driving model convergence and feature learning.

However, in coastal wetland environments, vegetation targets are often characterized by small object sizes, blurred boundaries, and mixed or mosaic distributions [31]. Relying solely on deep-layer features during final prediction may lead to insufficient utilization of detailed information contained in shallow features. Therefore, a second loss component is introduced: an auxiliary supervision applied to the auxiliary prediction generated from shallow features, also using cross-entropy. Shallow features retain higher-resolution textures and edge cues; enforcing additional supervision on them helps the model better capture small-scale structures and boundary details, and promotes multi-level feature consistency throughout the network.

The loss functions are defined as follows:

$$\begin{array}{c}{l}_{ce1}(Y,Y^{\prime })={\displaystyle \frac{{\sum }_{i=0}^H\sum _{j=0}^W[\sum _{l=1}^L{Y}_{i,j}^{\left(l\right)}\mathit{log}(Y_{i,j}^{\prime \left(l\right)})]}{H\times W}}\end{array}$$

(5)

$$\begin{array}{c}{l}_{ce2}(Y,upsample(Y^{″}))={\displaystyle \frac{{\sum }_{i=0}^H\sum _{j=0}^W[\sum _{l=1}^L{Y}_{i,j}^{\left(l\right)}\mathit{log}({upsample(Y}_{i,j}^{″\left(l\right)}))]}{H\times W}}\end{array}$$

(6)

where

• $Y^{\prime }$ denotes the prediction generated from the final decoding layer,
• $Y^{″}$ denotes the prediction obtained from deeper decoder features,
• $Y$ denotes the actual prediction result.
• $H,W$ denotes height and width
• $i,j$ denotes pixel row and column numbers

The total loss is defined as the sum of the main-branch and auxiliary-branch losses. This dual-supervision mechanism not only enhances the representational capability of the model but also improves its discriminative power and generalization performance when handling complex vegetation patterns in coastal wetlands:

$$l_{total}=l_{ce1}+l_{ce2}$$

(7)

2.3. Model Evaluation Indicators

To comprehensively evaluate the performance of the model in coastal vegetation classification, this study adopts several commonly used semantic segmentation and classification metrics, including Overall Accuracy (OA), Kappa coefficient (Kappa), Recall, and Mean Intersection over Union (mIoU) [32]. The calculation formulas for all indicators are shown in

Table 1. Formulas for the evaluation metrics used in model accuracy assessment.

	Formula
Accuracy	${\displaystyle \frac{TP+TN}{TP+FP+TN+FN}}$	(8)
Kappa	${\displaystyle \frac{P_0-P_e}{1-P_e}}$	(9)
$P_0$	(${\displaystyle \frac{TP+TN}{TN+TP+FP+FN}}$)	(10)
$P_e$	${\displaystyle \frac{(TP+FP)\times (TP+FN)+(FN+TN)\times (FP+TN)}{(TN+TP+FN+FP{)}^2}}$	(11)
Recall	${\displaystyle \frac{TP}{TP+FP}}$	(12)
Miou	$({\displaystyle \frac{TP}{TP+FP+FN}}+{\displaystyle \frac{TN}{TN+FP+FN}})/N$	(13)

. These indicators assess the model’s classification capability from multiple perspectives and provide a holistic evaluation of its effectiveness in identifying small vegetation patches, ambiguous boundaries, and mixed or mosaic distribution areas.

TP, TN, FP, FN can be expressed as:

TP: True sample size, the number of positive samples in the label and the number of positive samples in the predicted value.

TN: Number of true negative samples, number of negative samples in the label and number of negative samples in the predicted value.

FP: False Positive Sample Size, the number of negative samples in the label and the number of positive samples in the predicted value.

FN: False Negative Sample Size, number of positive samples in the label and negative samples in the predicted value.

2.4. Comparison Method

To thoroughly validate the effectiveness of the proposed method in coastal wetland vegetation classification, this study selects a set of representative semantic segmentation models for comparison. These models include classical convolution-based architectures, encoder–decoder structures, Transformer-based frameworks, and multi-scale feature fusion methods. Each model emphasizes different mechanisms of feature representation, enabling a comprehensive assessment of the proposed model’s advantages in detecting small vegetation targets, handling ambiguous boundaries, and extracting multi-scale features.

The comparison set includes four semantic segmentation models originally designed for natural images—UNet [33], DeepLabV3 [34], TransUNet [35], and SegFormer [36]—as well as two models specifically developed for remote sensing applications: D-LinkNet [37] and the boundary supervision-aided multiscale channel-wise cross-attention network (MCCA) [38].

3. Results

3.1. Coastal Wetland Vegetation Classification Mapping and Analysis

In this section, large-scale coastal wetland vegetation classification mapping was conducted using the proposed AFDFNet. The classification results are shown in Figure 6. As illustrated, AFDFNet maintains strong performance when applied to UAR high-resolution imagery in coastal wetland area. producing classification outputs that closely align with the spatial patterns observed in the RGB images.

Based on the mapping results, we further calculated the pixel proportion of each vegetation type within the study area and selected three representative species Spartina alterniflora, Reeds, and Suaeda salsa to evaluate the species-specific recognition accuracy of AFDFNet. The statistics are presented in Figure 7. The results indicate that the study area contains approximately 8% Reeds, 12% Spartina alterniflora, and 23% Suaeda salsa. The high proportion of Suaeda salsa reveals that the area supports a substantial amount of this native salt-tolerant species, which plays an important role in maintaining ecosystem function [39]. This suggests that the ecological condition of the region is generally healthy.

Meanwhile, the proportion of Reeds is lower than that of Spartina alterniflora. Given that Reeds is an indigenous dominant species commonly found in estuarine wetlands, whereas Spartina alterniflora is an invasive species, this distribution pattern indicates that the local vegetation ecosystem is being affected by the spread of Spartina alterniflora.

The proposed AFDFNet model exhibits excellent identification accuracy for all three vegetation types. The identification accuracy for each species exceeds 90%, further confirming that the AFDFNet model maintains robust performance even in complex wetland environments with coexisting species and mosaic spatial patterns.

3.2. Comparison and Analysis of Model Results

The performance of all models on the test set is summarized in

Table 2. Comparison of classification accuracies among different models.

	Accuracy	Kappa	Miou	Recall
unet	0.9032	0.7511	0.5422	0.6032
AFDFNet	0.9485	0.8274	0.6922	0.7371
transunet	0.9392	0.8566	0.6733	0.7066
D-linknet	0.8791	0.6894	0.4876	0.5431
deeplabv3	0.8922	0.6334	0.61	0.6822
MCCA	0.9086	0.7689	0.6499	0.7147
segformer	0.8527	0.6367	0.5368	0.6007

, and the visualized results are shown in Figure 8. As can be seen from the results in Figure 8b,f, the method proposed in this study still maintains excellent performance in the face of the complex symbiotic relationship between weak vegetation such as Suaeda salsa and Spartina alterniflora and Rees. The proposed AFDFNet exhibits outstanding performance, demonstrating clear advantages over other mainstream deep learning methods in terms of Overall Accuracy, Kappa coefficient, and Recall. In addition, TransUNet, which integrates Transformer and CNN architectures, also achieves competitive results and attains the best performance in Kappa. Specifically, compared with TransUNet, the proposed method improves Accuracy by 0.93%, mIoU by 1.89%, and Recall by 3.05%.

From Table 2, it can be observed that models using Vision Transformer (ViT)-based encoders struggle to maintain high performance when applied to ultra-high-resolution remote sensing imagery. The visual comparison further confirms this limitation. As shown in Figure 8f, SegFormer faces challenges in identifying regions where Reeds and Spartina alterniflora coexist. The model lacks sufficient detail representation, misclassifying clustered Spartina alterniflora as Reeds. This issue may arise because SegFormer employs a ViT encoder combined with a lightweight multi-layer perceptron (MLP) decoder. While this design enables efficient global context modeling and multi-scale feature aggregation, the MLP decoder lacks explicit spatial structural constraints and therefore does not recover fine-grained spatial details as effectively as convolution-based decoders [40]. Consequently, in complex boundaries or fine-textured regions such as the small clustered distribution of Spartina alterniflora the model produces blurred or overly smoothed predictions, causing small vegetation patches to be absorbed into surrounding Reeds areas.

A similar problem is observed with DeepLabV3 (Figure 8b). DeepLabV3 employs an Atrous Spatial Pyramid Pooling (ASPP) module to enlarge the receptive field through dilated convolutions; however, the sparse sampling nature of dilated convolution tends to weaken high-frequency texture information, such as the fine leaf structures of Spartina alterniflora [41]. In mixed-species zones, the suppression of high-frequency cues causes the model to rely more heavily on low-frequency regional shapes during classification, which ultimately leads to misclassification of Spartina alterniflora as Reeds.

4. Discussion

4.1. Evaluation of the AFDFNet Based on Ablation Experiments and Parameter Analysis

To further verify the contribution of each module within the proposed framework, this section employs the ResNet-50 backbone combined with a decoder in which the MFEM is removed as the baseline network. Based on this baseline, the following ablation experiments were designed:

(1) baseline + a: removing the frequency-domain interaction branch and replacing it with convolutional layers containing an equivalent number of parameters;

(2) baseline + b: removing the large-kernel convolution branch and replacing it with convolutional layers of similar parameter scale;

(3) baseline + a + b: removing the auxiliary loss used for deep-layer feature prediction.

The ablation results are presented in

Table 3. Results of the ablation experiments for the proposed model.

	Accuracy	Kappa	mIoU	Recall
baseline	0.9073	0.8578	0.6092	0.6322
Baseline + a	0.9351	0.7963	0.6331	0.7066
Baseline + b	0.9366	0.7752	0.6564	0.6891
Baseline + a + b	0.9431	0.8522	0.6884	0.7268

. It can be observed that each module contributes meaningfully to the overall performance improvement of AFDFNet in coastal vegetation classification. The substantial performance drop after removing the frequency-domain branch highlights the essential role of frequency-domain features. In coastal wetlands, vegetation types often exhibit weak spectral differences and highly interwoven, mosaic-like spatial patterns, making it difficult for models to distinguish species such as Reeds and Spartina alterniflora using spatial-domain cues alone [42]. Frequency-domain interactions enable explicit separation of high- and low-frequency components and strengthen both structural and boundary representations [43]. Therefore, removing this branch significantly weakens the model’s ability to capture fine-grained textures and delineate complex boundaries.

After removing the large-kernel convolution branch, the degradation in performance is even more pronounced: Accuracy decreases by 1.19%, Kappa by 5.22%, mIoU by 5.91%, and Recall by 4.8%. These results demonstrate that a large receptive field is indispensable for capturing the broad contextual information underlying wetland vegetation distribution. Coastal wetlands commonly exhibit interlaced vegetation patches, blurred boundaries, and small fragmented targets; traditional convolutional layers, limited by their narrow receptive field, struggle to model long-range dependencies [44]. In contrast, large-kernel convolutions substantially enhance global representation capacity, enabling the model to better recognize the macro-structures present in mixed-species regions [45]. Consequently, removing this branch results in the most severe performance deterioration.

When the auxiliary loss function is removed, the performance decline is comparatively modest—Accuracy decreases by 0.54%, Kappa by 0.44%, mIoU by 0.38%, and Recall by 1.03% but the impact remains noticeable. The shallow layer supervision primarily strengthens the sensitivity of early features to local structures and boundary details, while also improving training stability and convergence. Although its influence is not as dominant as the two core branches, the auxiliary supervision still improves detail prediction quality and provides measurable benefits in areas with complex boundaries.

In summary, the large-kernel convolution branch and the frequency-domain interaction branch are the key contributors to performance improvements. These two modules, respectively, enhance large-scale contextual modeling and frequency-domain feature representation. The auxiliary loss further promotes fine-detail learning and training efficiency. Working synergistically, these components allow the model to effectively handle spectral similarities, complex textures, and mosaic-like vegetation patterns in coastal wetlands, ultimately improving accuracy and robustness.

Model parameter size is another critical factor influencing practical deployment in high-resolution remote sensing classification tasks. UAV imagery often contains extremely large spatial dimensions and pixel densities, with a single image reaching thousands to tens of thousands of pixels. The inference speed of a model thus directly affects classification efficiency [46]. To evaluate the efficiency of the proposed framework during inference, we plotted the model parameter sizes against their corresponding mIoU values, as shown in Figure 9. The results show that AFDFNet achieves a favorable balance between accuracy and efficiency, outperforming other methods in both aspects. AFDFNet is capable of producing highly accurate coastal vegetation classification results with a relatively small number of parameters, demonstrating strong potential as a cost-effective solution for large-scale remote sensing applications.

4.2. Influence of Frequency-Domain Features on the Classification of Typical Coastal Vegetation

Building upon the ablation experiments, this section further investigates the impact of incorporating frequency-domain features on the classification accuracy of typical coastal vegetation types. Using the baseline + a model introduced in Section 4.1, we computed three evaluation metrics—Accuracy, IoU, and Recall—for three representative vegetation categories (Reeds, Spartina alterniflora, and Suaeda salsa). These results were then compared with the model performance obtained in Section 3.1. The comparison results are shown in Figure 10. Overall, the introduction of frequency-domain features leads to a performance pattern characterized as “overall improvement, local variation, and detail–structure trade-off.”

From the Accuracy metric, all three vegetation types exhibit performance gains: Reeds increases by 0.22%, Spartina alterniflora by 1.08%, and Suaeda salsa by 1.96%. This demonstrates that frequency-domain features enhance the model’s global discriminative ability, particularly in scenarios where spectral differences are subtle but texture differences are more pronounced. The larger improvements in Spartina alterniflora and Suaeda salsa suggest that frequency-domain cues are especially beneficial for vegetation types with distinctive texture patterns or small-scale patch structures.

Regarding the IoU metric, Reeds and Spartina alterniflora show improvements of 0.07% and 1.10%, respectively, indicating that frequency-domain features help the model better handle fine-grained textures and complex boundaries especially in the case of Spartina alterniflora, which typically exhibits dense, clustered growth patterns. However, Suaeda salsa shows a decrease of 0.57% in IoU. This suggests that while high-frequency enhancement strengthens detail representation, it may also amplify noise in small or boundary-blurred patches, causing boundary instability and reducing overall region-level consistency.

For Recall, all three vegetation types exhibit declines: Reeds drops by 0.03%, Spartina alterniflora by 1.31%, and Suaeda salsa—the most affected—by 2.90%. This indicates that frequency-domain enhancement, while improving edge sharpness and local texture discrimination, may slightly weaken the model’s sensitivity to the global structure of vegetation patches, leading to omission errors. This effect is most pronounced in Suaeda salsa, whose patches are typically small and have low spectral contrast. High-frequency amplification may overemphasize local variations, causing fragmentation of these patches and ultimately lowering Recall.

5. Conclusions

This study proposes a vegetation classification model tailored for high-resolution UAV imagery of coastal wetlands, where vegetation exhibits significant spectral mixing, complex symbiotic–mosaic patterns, and small target sizes. The model integrates spatial-domain and frequency-domain features through a dual-branch design: a large-kernel convolution branch enhances contextual modeling capability, while a frequency-domain interaction branch separates and fuses low-frequency structural information with high-frequency boundary details. In addition, shallow-layer auxiliary supervision is introduced to improve local detail learning and enhance training stability. The overall architecture maintains strong representational capacity while preserving a compact parameter size, making it suitable for real-world deployment in high-resolution remote sensing applications. Experimental results demonstrate that the proposed model achieves superior performance compared with baseline and state-of-the-art methods when identifying three typical coastal wetland vegetation types—Reeds, Spartina alterniflora, and Suaeda salsa.

Overall, the developed framework effectively addresses the challenges of weak spectral separability, widespread mixed growth patterns, and complex fine-scale structures in coastal wetland vegetation. It consistently outperforms mainstream segmentation models such as UNet, TransUNet, SegFormer, and DeepLabv3 across multiple key evaluation metrics, showing strong robustness and adaptability. The method provides a feasible and efficient technical pathway for UAV-based wetland ecological monitoring, invasive species detection, and landscape pattern analysis.

Future research may further explore adaptive frequency-domain fusion strategies, enhanced recall mechanisms for small objects, and multi-source or multi-temporal data integration to continuously improve the model’s generalization ability and practical value in complex wetland ecosystems.