In the task of semantic segmentation for remote sensing imagery, global contextual modeling is of paramount importance. Traditional Convolutional Neural Networks (CNNs) are intrinsically limited by their local receptive fields; while proficient in extracting fine textures, they struggle to capture the long-range, continuous linear structures characteristic of farmland shelterbelts, often leading to fragmented segmentation results. Conversely, although Transformer-based architectures can establish global dependencies via self-attention mechanisms, they suffer from quadratic computational complexity (O(N ² ) relative to sequence length. This poses a severe memory bottleneck and computational challenge when processing high-resolution remote sensing data. To address these limitations, we incorporate the Mamba-like Linear Attention (MLLA) operator (Gu and Dao, 2023) within the proposed Multi-Feature Fusion Block (MFFB). By leveraging Selective State Space Models (SSSM), Mamba achieves quasi-linear computational complexity (O(N)), enabling a global receptive field comparable to Transformers but with significantly reduced resource overhead. The integration of MLLA allows our model to perceive the extended trajectories of shelterbelts at a macro scale, ensuring geometric continuity of linear structures while simultaneously achieving a synergistic balance between global modeling capacity and inference efficiency.

While SSSM achieves efficient training through parallel scanning, recursive computation is less optimized on GPUs than matrix multiplication, resulting in performance bottlenecks. Han et al. (2024) demonstrated that SSSM is essentially a special variant of linear attention (LA) (Katharopoulos et al., 2020), which can be formulated as LA augmented with input and forget gates. On this basis, they developed the Mamba-like Linear Attention (MLLA) operator (Figure 3a). MLLA preserves global attention while reducing computational complexity to approximately O(N). In this study, we replace Transformer MHSA with MLLA and construct MLLAFormer (Figure 3b) to extract and fuse global contextual features from remote sensing imagery.

Convolutional modules are advantageous for local feature extraction, and their effectiveness has been validated in various computer vision tasks. To reduce computational complexity, this study adopts a combination of pointwise convolution (PConv) and depthwise convolution (DConv) as the local feature extraction operator. Furthermore, to enhance local feature representation and reduce potential redundancy, a channel attention operator is applied after convolution, with its computation defined as Equation 2. Φ X = σ W l   ω X (2)

In Equation 2, X denotes the input feature map, ω represents global average pooling (GAP), W l is the weight parameter of a linear transformation, and σ denotes the Sigmoid activation function. The convolutional block with this operator is illustrated in Figure 3c. Lin et al. (2023) showed that the architectural design of Transformers is one of the key factors for their success. To leverage this design while maintaining structural consistency among feature extraction operators, this study replaces Transformer MHSA with the proposed convolutional block to construct ConvFormer (Figure 3d), which is employed to extract local features from remote sensing imagery.

The multi-layer perceptron (MLP) in Transformers tends to focus on low-frequency features in the spatial domain (Rahaman et al., 2019), limiting its ability to capture fine textures and sharp edges. This constraint is unfavorable for reconstructing high-frequency details in the super-resolution process and for preserving fine object boundaries in semantic segmentation results. Fourier feature transformation, by contrast, is effective in capturing high-frequency details (Feng and Liu, 2025). Therefore, in this study, we employ the fast Fourier transform (FFT) to convert inputs from the spatial domain to the frequency domain, where an MLP directly extracts image features. On this basis, we construct FFTFormer (Figure 3e) to capture frequency-domain features from remote sensing imagery.

MLLAFormer, ConvFormer, and FFTFormer extract heterogeneous features from remote sensing imagery in the global, local, and frequency domains, respectively. Across different layers of the semantic segmentation and super-resolution networks, these three operators should contribute unequally; thus, naive feature concatenation or summation cannot meet this requirement. To adaptively balance their contributions, we propose a Spatially Gated Fusion Mechanism (SGFM), which computes fusion weights in a spatially aware manner. The computation of SGFM is defined in Equations 3, 4. G = σ ψ 1 X 1 , X 2 (3) Y = G ⊙ X 1 + 1 − G ⊙ X 2 (4)

In Equations 3, 4, X 1 , X 2 ∈ R c × h × w denotes the feature maps extracted by different operators; , indicates channel-wise concatenation; ⊙ denotes element-wise (Hadamard) multiplication; ψ 1 is a pointwise convolution producing a single-channel output; G ∈ R 1 × h × w is the spatial gating map; Y ∈ R c × h × w denotes the fusion result of X 1 and X 2   in the spatial domain, where c, h and w represent the number of channels, height, and width of the feature map, respectively. The structure of the Spatially Gated Fusion Mechanism (SGFM) is illustrated in Figure 4a. The gating map G is generated jointly from X 1 , X 2   and their features are fused through G to produce Y. During backpropagation, the gradients are propagated via G into the corresponding feature extraction operators of both X 1 and X 2   . As a result, each operator not only reduces its own error but also contributes to minimizing the errors of the others, making SGFM function as a lightweight cross-attention mechanism.

Different types of features are extracted using MLLAFormer, ConvFormer, and FFTFormer,and subsequently fused through the Spatially Gated Fusion Mechanism (SGFM) to construct the Multi-Feature Fusion Block (MFFB), whose structure is illustrated in Figure 4b. The MFFB consists of local and global sub-layers. In the local sub-layer, the input feature map F i n ∈ R c × h × w is first processed by two parallel branches, ConvFormer and FFTFormer, to extract local and frequency-domain features. These features are then fused by SGFM to generate the intermediate result Y. To reuse the original information, a third parallel branch is introduced in the local sub-layer, which scales the input feature map with a learnable parameter γ ∈ R c and further fuses it with Y, as formulated in Equation 5. F o u t = γ F i n + Y (5)

In Equation 5, denotes the output of the local sub-layer. In the global sub-layer of MFFB, MLLAFormer replaces ConvFormer to perform global feature extraction. FFTFormer is applied in both the local and global sub-layers and, together with SGFM, guides the extraction of frequency-aware local and global features. Consequently, MFFB captures local and global information in both the spatial and frequency domains, while the gated fusion mechanism balances the relative importance of heterogeneous features. Building on the proposed MFFB, we construct remote sensing super-resolution and semantic segmentation models.

We construct the remote sensing super-resolution network using the proposed MFFB as the core feature extraction unit. Following the progressive upscaling paradigm commonly adopted in prior work (Qiao et al., 2024) (Figure 5a), a convolutional stem first encodes basic features ( F b ∈ R d × h × w ) from the low-resolution (LR) image with d channels; then L feature extraction layers (FELs) are stacked, each implemented by repeatedly applying the MFFB D times to deepen representation capacity while keeping both the input and output channel dimensions at d.

We stack L carefully designed feature extraction layers (FELs). After the Lth FEL, a convolutional layer adjusts the channel dimension of the output feature map to match the basic feature F b , yielding the high-level feature F h . A long skip connection then fuses F b and F h via element-wise addition to obtain F h , from which an upsampler reconstructs the high-resolution (HR) image. Following this architecture, we instantiate each FEL with the proposed MFFB and repeat MFFB D times within the FEL to deepen representation capacity; a subsequent convolutional layer and layer normalization are appended to stabilize training, and a local skip connection element-wise fuses the FEL input and output for feature reuse. The input and output channel width of every FEL is fixed at d. Consequently, (d,L,D) serve as hyperparameters controlling the complexity of the super-resolution model.

Interpolation, pixel shuffle, and transposed convolution are commonly used upsamplers in super-resolution networks. While interpolation and pixel shuffle are static upsampling methods without learnable parameters (Guo et al., 2024), relying on a single strategy limits the ability to capture diverse features. In this study, we design an upsampler that combines pixel shuffle and transposed convolution, as illustrated in Figure 5b. The upsampler contains two parallel branches: one branch applies a transposed convolution with stride R and kernel size R × R to upsample by a factor of R, while the other branch performs pixel shuffle upsampling. To endow the pixel shuffle branch with learning capability, a convolutional layer is first introduced to expand the input feature channels to dR², which are then rearranged spatially by pixel shuffle into blocks of size d × R × R, thus achieving R-fold upsampling. The outputs of the two branches are adaptively fused by SGFM, and a pointwise convolution with three output channels is applied to reconstruct the HR image. Using this upsampler, the final super-resolution model is constructed, enabling LR remote sensing images to be super-resolved into HR images of size R times larger.

Building on the multi-domain feature extraction capability of MFFB in the global, local, and frequency domains, we construct the semantic segmentation model (Figure 6a). The encoder consists of four stages. In each stage, a convolutional module (Figure 6b) is first applied for basic feature extraction, adjusting the channel dimension of the feature map to   D i , i ∈ 1 , 2 , 3 , 4 denotes the stage index, while strided convolution is used for spatial downsampling. This is followed by stacking L i MFFB modules. Here, L i and   D i are treated as encoder hyperparameters to control the network width (output channel dimension) and depth (number of layers), respectively. Each stage generates feature maps   F i ϵ R D i × H / 2 i + 1 × W / 2 i + 1 at different spatial scales, where H and W denote the height and width of the input image. These multi-scale feature maps F i form a feature pyramid, which serves as the input to the decoder.

The decoder first aggregates multi-level encoder features ( F 2 , F 3 and   F 4 ). Three parallel branches upsample these features to a common spatial resolution. In each branch, a pointwise convolution adjusts the channel dimension of F i from   D i to C, followed by bilinear interpolation to obtain F u p i ∈ R C × H / 8 × W / 8 . The parameter C serves as a decoder hyperparameter to balance modeling capacity and computational cost. The three upsampled features are then concatenated along the channel dimension and fused via a pointwise convolution to perform feature aggregation and channel alignment, as formulated in Equation 6. F = ψ N F u p 2 , F u p 3 , F u p 4 (6)

In Equation 6, F ∈ R N × H / 8 × W / 8   denotes the aggregated feature map and N represents the number of channels. N is treated as another decoder hyperparameter, serving the same role as C in balancing modeling capacity and computational complexity.

Since the three parallel branches only involve pointwise convolution and upsampling, they lack the ability to model multi-scale contextual environments. Incorporating multi-scale context modeling of the input imagery can enhance cross-scale semantic representation, mitigate intra-class scale variation, and improve inter-class discrimination. To this end, we design a Multi-Scale Context Extraction (MSCE) module (Figure 6c) to strengthen the decoder’s multi-scale contextual modeling capacity. The input to MSCE is the aggregated feature map F. The module consists of three branches: in Branch 1, pixel unshuffle (PU) rearranges non-overlapping 4 × 4 spatial blocks of F into the channel dimension to produce F s 1 ∈ R 16 N × H ⁄ 32 × W ⁄ 32 , effectively encoding a receptive field of 4 × 4 without information loss. In Branch 2, a 3 × 3 convolution with stride 2 is applied to F, followed by PU that rearranges non-overlapping 2 × 2 blocks into the channel dimension, generate F s 2 ∈ R 8 N × H / 32 × W / 32 with an effective receptive field of 5 × 5. In Branch 3, a 5 × 5 convolution with stride 4 is used to capture larger-scale contextual features, producing F s 4 ∈ R 4 N × H ⁄ 32 × W ⁄ 32 . These branches provide receptive fields of different scales and thus capture contextual information at multiple levels. To balance contributions across scales while reducing computational cost, pointwise convolution is first applied to adjust the channel dimension of F s 1 , F s 2 and F s 4 to N, after which the three are fused. Although MSCE extracts multi-scale contextual features, global context is still absent. Therefore, we further apply MLLAFormer to the concatenated multi-scale features to integrate global context, with the overall fusion process defined in Equation 7. F s = MLLAFormer ψ N F s 1 , ψ N F s 2 , ψ N F s 4 (7)

In Equation 7, denotes the multi-scale contextual features enhanced with global information. These features are further upsampled to match the spatial resolution of F, and a pointwise convolution is applied to adjust the channel dimension, as formulated in Equation 8. F ′ = ψ N U F s (8)

In Equation 8, F ′ ∈ R N × H ⁄ 8 × W ⁄ 8 denotes the multi-scale contextual features generated by MSCE, and U represents the upsampling operation. The overall computation of MSCE can be concisely expressed as Equation 9. F ′ = MSCE F (9)

Compared with F, F ′ incorporates multi-scale and global contextual information, which provides crucial auxiliary cues for foreground classification (Terven et al., 2023). Since F u p 2 , F u p 3 and F u p 4 are suited to segment foreground targets at different scales, we fuse F ′ into each of these feature maps. To retain fusion flexibility, we employ SGFM; taking the fusion with F u p 2 as an example, the process is formulated in Equation 10. F u p 2 ′ = SGFM F u p 2 , F ′ (10)

In Equation 10, F u p 2 ′ denotes the fused feature. By the same procedure, F ′ is fused with F u p 3 and F u p 4 to obtain F u p 3 ′ and F u p 4 ′ , respectively.

MSCE takes F as input and fuses it to produce F ′ , a multi-scale contextual feature with global information. To further enhance the diversity of contextual features, following the approach of DeepLab v3+, global average pooling (GAP) is employed to extract the global feature F g = GAP F , which is then utilized to assist semantic segmentation. Simultaneously, to leverage F potential for semantic segmentation, it undergoes a transformation via pointwise convolution, yielding F c = ψ N F , which is then applied for semantic segmentation. Multiple feature types are cascaded, and an MLP composed of two pointwise convolutions predicts the semantic segmentation mask M for the input image. This process is expressed as Equation 11: M = U MLP F u p 2 ′ , F u p 3 ′ , F u p 4 ′ , F ′ , F g , F c (11)

The equation employs an upsampling function U to upscale the segmentation mask predicted by the MLP to the same size as the input image I. The structure of the decoder and the entire remote sensing image semantic segmentation network is shown in Figure 6a.

Training and testing experiments for super-resolution models and semantic segmentation models were conducted on multiple computers equipped with Intel Xeon Silver 4314 processors, 128 GB of memory, Nvidia L20 compute cards, Windows Server 2022 operating systems, and CUDA 11.8 acceleration libraries. Based on PyTorch 2.7, all models were implemented using Python, alongside the corresponding computational programs for the technical framework depicted in Figure 2. The farmland forest network dataset, comprising research area imagery and annotation results, was divided into five equal parts. Four parts were randomly selected as training data, with the remaining part reserved as test data.

The super-resolution model was trained using the AdamW optimizer with a learning rate of 2 × 10⁻⁴. Training samples were dynamically generated. Taking an upsampling factor R = 2 as an example: first, randomly crop image patches with width and height ranging from 128 to 512 pixels from the training images. Then, resize these patches to 128 × 128 pixels as high-resolution (HR) images. Perform bilinear interpolation downsampling on the HR images to generate low-resolution (LR) images with a resolution of 64 × 64 pixels. Use LR and HR images as sample pairs to train the super-resolution model. Existing super-resolution models (Ariav and Cohen, 2022) commonly employ L1 as the loss function. To enhance the structural accuracy of local super-resolution reconstruction, we further incorporate the SSIM loss L _SSIM (Wang et al., 2004).The loss function Lsup for our super-resolution model is defined as shown in Equation 12: L sup = λ 1 H R − H R ¯ 1 + λ 2 L S S I M H R , H R ¯ (12)

In the equation, λ 1 and λ 2 are the proportional adjustment coefficients for the L₁ and L_SSIM losses, respectively, and H R ¯ is the super-resolution reconstructed image.

The semantic segmentation model was trained using the AdamW optimizer with a learning rate of 10⁻⁵ and input images at a resolution of 512 × 512 pixels. Randomly cropped image patches and their corresponding semantic annotations from the training dataset were used as training samples, with data augmentation applied through random flipping and color dithering. To address the pixel imbalance across three land cover classes in the study area imagery, the loss function L for the semantic segmentation model was defined using both the commonly applied cross-entropy loss (L _CE ) and the Dice loss (L _Dice ), which is more robust to sample imbalance, as shown in Equations 13–15. L = α L C E + β L D i c e (13) L C E = − 1 H W M g t ⁡ log M (14) L D i c e = 1 − 2 M g t M M g t 2 + M 2 (15)

In the equation, M _gt represents the ground truth semantic labels corresponding to the input image, with α = 0.4 and β = 0.6 as proportional adjustment coefficients. Using arg min L as the objective, the semantic segmentation model is trained and optimized until the loss converges to a stable value.

The performance of semantic segmentation models is evaluated using precision, recall, intersection over union (IoU), pixel accuracy (PA), and mean intersection over union (mIoU) (He et al., 2024). The definitions of each metric are given in Equations 16–20: p r e c i s i o n i = T P i T P i + F P i (16) R e c a l l i = T P i T P i + F N i (17) IoU i = T P i T P i + F P i + F N i (18) P A = ∑ i = 1 3 T P i N t (19) mIoU = 1 3 ∑ i = 1 3 IoU i (20)

In the equation, T P i denotes the number of pixels correctly predicted as category i, F P i denotes the number of pixels incorrectly predicted as category i, F N i denotes the number of pixels in category i incorrectly predicted as other categories, and N t denotes the total number of pixels in the test set samples. The performance of super-resolution models is evaluated using peak signal-to-noise ratio (PSNR) (Tanchenko, 2014) and structural similarity index (SSIM).

To rigorously quantify the geometric fidelity and boundary integrity of the vectorized shelterbelts, we introduced the “Shelterbelt Closure Index” as a key evaluation metric. According to the Chinese National Standard Design Specifications for Farmland Shelterbelt Engineering (GB/T 50817-2013), the closure index of shelterbelts surrounding farmland parcels in severely wind-sand affected areas should be ≥0.75 to ensure effective protection. The calculation methodology is implemented as follows: First, a buffer zone with a radius of 150 m is generated for each individual field polygon to identify the corresponding protective shelterbelts Second, the “protection angle” is defined as the maximum geometric angle formed by connecting the geometric centre of the field polygon to the boundary extremities of a single shelterbelt (as illustrated in Figure 7) Finally, the closure rate is calculated as the ratio of the cumulative sum of all unique protection angles within the buffer zone to 360°, with a value ≥ 0.75$ serving as the threshold for meeting the geometric standard.

To validate the effectiveness of the super-resolution model proposed in this study, we compared it with several state-of-the-art super-resolution methods under a 4x upsampling task. These included the traditional bicubic interpolation baseline method, the Transformer-based super-resolution model TransENet (Lei et al., 2022), and the Mamba-based models MambaIR (Guo et al., 2024) and FreMamba (Xiao et al., 2024). The hyperparameters for our super-resolution model were set to d = 96, D = 6, and L = 6. All methods were trained and tested on the farmland forest network dataset. The comparison models employed the same training methodology as their original literature. Results are presented in Table 1.

As shown in Table 1, the super-resolution model developed in this study demonstrates the highest performance, outperforming the comparison methods in both PSNR and SSIM metrics. This approach is effective for the upsampling task on remote sensing images of farmland shelterbelts. Compared to TransENet, our model achieves a 0.44 dB higher PSNR and a 1.58 percentage point improvement in SSIM, while requiring only 36% of TransENet’s parameters, indicating superior parameter efficiency. Our Multi-Feature Fusion Block (MFFB) simultaneously extracts features from both spatial and frequency domains, offering more diverse feature extraction capabilities than the Transformer architecture. Compared to FreMamba and MambaIR, our model also demonstrates significant advantages. In terms of PSNR, it outperforms the former by 0.47 dB and the latter by 1.63 dB. For SSIM, it achieves improvements of 1.76 and 3.34 percentage points, respectively. Despite having 18% more parameters than FreMamba, our model achieves better performance gains, reflecting a favorable performance-efficiency balance. All deep learning-based models substantially outperformed the traditional bicubic interpolation benchmark. Compared to this baseline, the proposed model achieved a 6.10 dB improvement in PSNR and a 7.24 percentage point increase in SSIM, confirming the effectiveness of feature-reconstruction-based upscaling methods. Overall, the proposed super-resolution model demonstrates applicability for upscaling remote sensing imagery of farmland shelterbelts.

To analyze the impact of introducing LSSIM loss during the design and training of MFFB and Upsampler on the super-resolution model in this study, further ablation experiments were conducted. Model 1 replaces MFFB’s MLLAFormer and FFTFormer with ConvFormer, replaces SGFM with element-wise addition, and replaces the upsampler with PixelShuffle. Training uses only L1 loss, meaning MLLAFormer, FFTFormer, SGFM, Upsampler, and L _SSIM are disabled in Model 1. Building upon Model 1, subsequent processing steps were progressively enabled to construct Models 2 through 6. All models performed 4x upsampling and underwent identical training and testing on the farmland forest network dataset, with results presented in Table 2.

As shown in Table 2, PSNR and SSIM gradually increase with the introduction of different modules and processing steps, indicating that MFFB, Upsampler, and L _SSIM all enhance the super-resolution performance of remote sensing images of farmland shelterbelts. Model 1 (baseline model) achieves a PSNR of 31.34 dB without enabling MLLAFormer processing, comparable to TransENet in Table 1. At this stage, MFFB consists solely of the ConvFormer base module, demonstrating the effectiveness of the ConvFormer architecture design. Compared to Model 1, Model 2 achieved a 0.1 dB increase in PSNR and a 1.53 percentage point improvement in SSIM, demonstrating MLLAFormer’s ability to enhance super-resolution performance. Unlike ConvFormer, MLLAFormer replaces convolutions with MLLA, whose global modeling capability strengthens the spatial dependencies of features and improves the spatial continuity reconstruction in super-resolution. Compared to Model 2, Model 3 achieves a PSNR improvement of 0.15 dB and an SSIM increase of 0.57 percentage points. FFTFormer extracts features in the frequency domain, strengthening image texture detail and edge recovery capabilities. Model 4 enables SGFM on top of Model 3, resulting in marginal PSNR and SSIM gains of 0.05 dB and 0.13 percentage points, respectively. Model 5 achieves a 0.06 dB PSNR improvement after enabling the Upsampler. By combining transposed convolution with pixel rearrangement, the Upsampler enhances the model’s upscaling capability. Model 6 introduces L _SSIM on top of Model 5, achieving 0.05 dB and 0.7 percentage point improvements in PSNR and SSIM, respectively. L _SSIM better constrains the model’s reconstruction quality of local image structures, thereby improving overall image reconstruction. Compared to Model 1, Model 6 achieved 0.41 dB and 2.99 percentage point improvements in PSNR and SSIM respectively, indicating the effectiveness of all processing steps. The modular design in this study enhances the super-resolution performance of remote sensing imagery of farmland shelterbelts. Model 6 is adopted as the super-resolution model for subsequent experimental analysis.

To investigate the impact of different upscaling strategies on super-resolution model performance, we retrained Model 6 from Table 2 with an upscaling factor of 2. Based on this, we compared three upscaling approaches: 2× upscaling using the newly trained model, 4× upscaling achieved through a two-step staged super-resolution process with the newly trained model, and direct 4x upscaling using the original Model 6. The results are presented in Table 3.

Table 3 demonstrates that the upsampling strategy significantly impacts the super-resolution reconstruction performance of remote sensing images of farmland shelterbelts. Compared to 2× upsampling, 4× upsampling yields poorer reconstruction results for super-resolution images. 2× upsampling only requires expanding a single LR pixel into a 2 × 2 HR pixel block, making it easier for the model to learn local high-frequency details and low-frequency structures. Compared to 2× upsampling, direct 4x upsampling resulted in PSNR and SSIM decreases of 9.01 dB and 13.04 percentage points, respectively. 4× upsampling requires predicting a 4 × 4 HR image block from a single LR pixel and its surrounding blurred structures, making it challenging for the model to accurately infer fine local details with limited contextual features. The staged upsampling process outperforms direct 4x upsampling, achieving a 2.85 dB improvement in PSNR and a 1.48 percentage point increase in SSIM. First, 2× upsampling generates high-quality intermediate-resolution images, providing reliable input features for the second-stage MFFB. This enables the model to enhance details and sharpen images based on partially optimized data. This strategy mitigates the challenges of high-magnification upsampling while leveraging the progressive learning advantages of the MFFB feature extraction module. In the super-resolution task for remote sensing images of farmland shelterbelts, the upscaling strategy significantly impacts image reconstruction quality. Under identical upscaling factors, staged upscaling outperforms single-stage approaches. All subsequent experiments involving 4× upscaling employ the two-stage upscaling process.

To systematically evaluate the effectiveness of the semantic segmentation model developed in this study for farmland forest network segmentation tasks, five representative semantic segmentation models were selected for comparative analysis. These include RS³ Mamba (Ma et al., 2024a) based on Mamba, DC-Swin (Wang et al., 2022) and GASOT-Net (Zhang et al., 2024) based on Swin Transformer, SSRS (Ma et al., 2024b) fine-tuned on SAM (Kirillov et al., 2023), and the convolutional-based DeepLabV3+ and UNet. The hyperparameters for the model under study were set as follows: L1 = 2, L2 = 4, L3 = 8, L4 = 4, D1 = 64, D2 = 128, D3 = 256, D4 = 512, C = 320, N = 96. All models were trained and tested using the farmland forest network dataset, with inputs being two-stage quadruple-upsampled images. Results are presented in Table 4.

As shown in Table 4, the semantic segmentation model developed in this study outperforms the comparison models across all metrics. The segmentation accuracy for cultivated land reached 96.42% with an IoU of 0.9513, while the accuracy for shelterbelt segmentation was 82.83% with an IoU of 0.7403. The model achieved a pixel accuracy (PA) of 94.89% and a mean IoU (mIoU) of 0.8345. Compared with RS³ Mamba, our model improved the IoU for cultivated land segmentation by 1.31%, increased the pixel recall for shelterbelts by 3.19 percentage points, enhanced PA by 1.32 percentage points, and boosted mIoU by 1.14%. The MFFB module in this model employs FFTFormer to capture high-frequency textures like canopy edges, mitigating confusion between forest strips and cultivated crops at marginal locations. MLLAFormer establishes long-range spatial dependencies, reducing fragmentation in elongated forest strips. Compared to the Transformer-based DC-Swin, our method achieves advantages of 1.09% in mIoU and 0.34% in PA, indicating that the multi-feature fusion mechanism outperforms the single attention paradigm. In comparison with GASOT-Net, which is also based on Swin Transformer, our method achieves a marginal improvement in mIoU 0.08%, while significantly increasing the segmentation accuracy of shelterbelts by 0.72 percentage points, further demonstrating the effectiveness of multi-domain feature fusion for slender objects. Compared to SSRS, our model achieves a 1.32% improvement in mIoU. Agairbelts respectively, along with a 4.13% increase in mIoU and a 3.79 percentage point rise in PA. When benchmarked against UNet, our model achieves a 26.20% improvement in mIoU. The MFFB feature fusion in this model, achieved through a triadic approach combining global, local, and frequency-domain features, addresses the challenge of capturing slender forest strips and objects with similar textures—a limitatinst DeepLabV3+, our model demonstrates 3.70% and 3.65% higher IoU for cropland and shelteon when relying solely on spatial domain convolutions. All methods demonstrated higher segmentation accuracy for farmland than for shelterbelts, indicating that precise segmentation of elongated objects remains a challenge. Overall, the proposed semantic segmentation model is suitable for farmland forest network segmentation.

To further verify the generalization capability of the proposed method, this study conducted additional comparisons with various baseline methods on the LOVEDA dataset. The experiment quantified the segmentation accuracy of five foreground categories, with the results of each category presented in terms of the Intersection over Union (IoU), as detailed in Table 5.

The results in Table 5 demonstrate that the proposed method achieves the highest mIoU on the public LOVEDA dataset, confirming its strong generalization capability. The model performs particularly well on structurally distinct categories such as “building,” “road,” and “river,” indicating that its integrated frequency-domain and global features offer significant advantages in extracting generalizable edge and structural information. Compared to other models, the proposed method maintains a leading performance on datasets with more complex scenes and greater object diversity, proving that its design is not only tailored to the specific task of farmland shelterbelt extraction but also exhibits broad applicability.

Building upon the evaluation of model accuracy, to comprehensively analyze the feasibility and efficiency advantages of the proposed method in practical applications, we further compared the computational complexity and inference speed of various semantic segmentation models. Table 6 presents the performance metrics of the proposed method and the comparative models in terms of parameter count (Params), computational cost (GMACs), and average inference time per image (Latency). All tests were conducted under identical hardware conditions, with an input image resolution of 512 × 512 pixels and a batch size of 1, reflecting the operational efficiency of the models in real deployment scenarios. The experimental results are presented in Table 6.

In terms of computational efficiency and deployment performance, the proposed method achieves a favorable balance between model complexity and inference speed. Although its parameter count is higher than that of lightweight models such as RS³ Mamba, it is significantly lower than the computationally intensive GASOT-Net. In terms of computational cost (GMACs), the proposed method is only 45.5% of GASOT-Net and lower than UNet, indicating that its multi-feature fusion mechanism enhances representational capacity without introducing excessive computational overhead. Moreover, The inference latency of 29.2 ms indicates that the MFFB module significantly enhances feature representation without introducing exponential computational overhead. Overall, while maintaining high segmentation accuracy, the proposed method exhibits relatively superior computational efficiency and real-time processing capability, making it suitable for remote sensing extraction tasks of farmland shelterbelts that require both efficiency and precision.

To analyse the contribution of MFFB’s core submodules—MLLAFormer, FFTFormer, and SGFM—to semantic segmentation, four variant models were constructed. Baseline Model A disables MLLAFormer, FFTFormer, and SGFM, replacing them within MFFB with ConvFormer for MLLAFormer and FFTFormer, and element-wise addition for SGFM. Model B enables MLLAFormer on top of Model A, Model C enables FFTFormer on top of Model B, and Model D further enables SGFM. On the farmland forest network dataset, each variant model was trained and tested using identical methods, with results shown in Table 7.

As shown in Table 7, the semantic segmentation performance of the model gradually improves with the progressive introduction of submodules in MFFB. Compared to Model A, Model B achieves a 0.70% increase in cultivated land segmentation IoU, a 0.90 percentage point improvement in precision, and a 1.04% rise in Farmland Shelterbelts segmentation IoU. This demonstrates that MLLAFormer’s global context modeling capability enhances segmentation for both large-scale and elongated objects. Compared to Model B, Model C achieves a 0.53 percentage point increase in cultivated land segmentation accuracy, a 1.61 percentage point increase in protective forest belt segmentation accuracy, a 0.45% improvement in IoU for cultivated land, a 2.80% improvement in IoU for protective forest belts, and a 1.76% increase in mIoU. This demonstrates that FFTFormer’s frequency domain feature extraction plays a crucial role in segmenting elongated objects. Frequency-domain feature extraction enhances the clarity of elongated object boundary features, improving the distinguishability between different object boundaries. Compared to Model C, Model D achieves a 1.25 percentage point increase in Farmland Shelterbelts segmentation accuracy and a 1.71% improvement in mIoU, indicating that SGFM contributes to enhancing semantic segmentation accuracy. SGFM’s spatial gated weighting coordinates the synergistic optimization of local, global, and frequency-domain spatial features, positively enhancing the model’s semantic segmentation performance. Compared to Model A, Model D achieved a 1.56 percentage point increase in cultivated land segmentation accuracy and a 1.87 percentage point increase in IoU, while Farmland Shelterbelts segmentation accuracy and IoU improved by 3.64 and 4.08 percentage points, respectively. while mIoU increased by 5.14%, validating the effectiveness of the MFFB module design. Overall, the coupled design of multi-feature fusion effectively improved the semantic segmentation accuracy of farmland shelterbelts. Model D is adopted as the final semantic segmentation model for this study.

Further quantitative evaluation of the effectiveness of super-resolution preprocessing for semantic segmentation of farmland shelterbelts. Using the same semantic segmentation model, three sets of comparative experiments were conducted: Experiment I served as the control group, directly inputting images at native resolution; Experiment II used images upsampled via 2× super-resolution; Experiment III employed images upsampled via a two-stage 4× super-resolution process. The semantic segmentation accuracy was calculated for each group, with results presented in Table8.

As shown in Table 8, super-resolution preprocessing significantly enhances semantic segmentation performance. Compared to the control group, 4× upsampling improved segmentation accuracy for farmland and shelterbelts by 0.91 and 4.32 percentage points respectively, increased IoU by 1.85% and 6.61%, and boosted mIoU for semantic segmentation by 2.47%. Super-resolution preprocessing enhances segmentation for both farmland and shelterbelts, particularly improving the segmentation accuracy of shelterbelts. Super-resolution reconstruction amplifies high-frequency components in images, thereby improving the texture discrimination capability of semantic segmentation models. Simultaneously, super-resolution preprocessing enhances the width and edge detail information of narrow shelterbelts, effectively alleviating the challenge of segmenting elongated objects faced by semantic segmentation models. This study also conducted segmentation experiments with 8× upsampling, but encountered rapid performance degradation. This was attributed to artifacts appearing in the significantly upsampled images, whose style differed markedly from the training samples, causing a substantial decline in semantic segmentation model performance. Therefore, these results are not reported. Overall, super-resolution preprocessing at certain magnifications significantly improves semantic segmentation accuracy, particularly for challenging narrow Farmland Shelterbeltss.