Assessing GAN Super-Resolution in Grasslands: The Role of Spatial Heterogeneity and Textural Complexity

Highlights

What are the main findings?

• Landscape heterogeneity (vegetation condition, structure, and texture) strongly controls super-resolution performance, with SRGAN degrading under high complexity while ESRGAN remains robust.
• ESRGAN consistently outperforms SRGAN and bicubic interpolation across intra-sensor, cross-sensor, and generalization scenarios, particularly under domain transfer conditions

What are the implications of the main findings?

• Incorporating ecological heterogeneity into model evaluation is essential for reliable deployment of deep learning super-resolution in remote sensing.
• The demonstrated robustness of ESRGAN positions it as a scalable solution for multi-sensor data fusion in heterogeneous ecosystems.

Abstract

High-resolution imagery is essential for monitoring heterogeneous grassland ecosystems, yet the performance of generative adversarial network (GAN) super-resolution under varying landscape heterogeneity and operational application scenarios remains unclear. This study presents a landscape-aware evaluation of super-resolution methods in semi-arid savanna grasslands of the Edwards Plateau (Texas, USA) using paired multispectral imagery from PlanetScope (3 m) and unmanned aerial vehicle (UAV) platforms (0.03 m). Two GAN models, SRGAN and ESRGAN, were compared with a bicubic interpolation baseline. Image tiles were systematically stratified along ecologically relevant gradients of vegetation condition (NDVI quartiles), spatial structure (woody patch-based clusters), and textural complexity (GLCM entropy quartiles). Model performance was evaluated across three operational frameworks: intra-sensor downscaling, cross-sensor downscaling, and intra-to-cross generalization. Reconstruction fidelity was quantified using peak signal-to-noise ratio (PSNR) and structural similarity index (SSIM), complemented by variability analysis to assess performance stability. Landscape heterogeneity strongly influenced downscaling outcomes. SRGAN performance declined in areas with dense vegetation, aggregated woody structure, and high-entropy textures, with large variability under cross-sensor and generalization scenarios. In contrast, ESRGAN demonstrated consistently robust performance across landscape gradients, whereas bicubic interpolation performed well only under intra-sensor conditions and drastically degraded under sensor transfer. These results demonstrate that vegetation condition, structural heterogeneity, and sensor-transfer scenarios jointly constrain super-resolution performance. Rather than serving as a model comparison exercise, this study emphasizes a landscape-aware framework for understanding how ecological heterogeneity and operational domain shifts jointly shape super-resolution behavior in grassland ecosystems, providing guidance for more reliable applications of deep learning-based remote sensing methods.

1. Introduction

Monitoring grassland ecosystems requires spatially explicit, high-resolution information capable of capturing fine-scale vegetation dynamics, detecting early signals of degradation, and supporting precision management. Recent advances in remote sensing, especially satellite platforms combined with unmanned aerial vehicles (UAVs), have expanded monitoring capacity across scales and seasons. However, heterogeneous grassland mosaics, characterized by diffuse edges, bare patches, and micro-topography, remain challenging for conventional approaches that rely on coarse pixels or global accuracy metrics [1,2]. Landscape ecology research further emphasizes that spatial heterogeneity and patch structure strongly mediate the information content available to remote sensing analyses, reinforcing the need for evaluation protocols that reflect ecological pattern and scale rather than relying solely on global averages [3].

Grasslands cover approximately 40% of Earth’s ice-free land surface and 69% of agricultural land, contributing nearly 20% of terrestrial productivity [4,5]. These ecosystems sustain millions of livelihoods through livestock production while providing critical ecosystem services, including carbon sequestration, biodiversity conservation, and soil stabilization [1,2,6,7]. Despite their ecological and socio-economic significance, grasslands are among the most threatened and least protected biomes worldwide, facing degradation driven by land conversion, overgrazing, and climate change; therefore, their protection and management is a pressing issue [8,9,10,11,12,13,14]

Satellite sensors such as PlanetScope (3 m resolution) offer frequent revisit times conducive to monitoring dynamic landscapes. However, ref. [15] show that this resolution may be still insufficient for heterogeneous rangelands, where scattered shrubs, bare soil patches, and micro-topographic variations often fall below the pixel scale, leading to misclassification of degradation stages. Similarly, sentinel-2 imagery (10 m resolution) improves biomass and leaf area index estimates but struggles to characterize fragmented parcels and narrow management strips due to mixed pixels and edge effects [1]. In contrast, UAV multispectral imagery complements satellite data by providing superior spatial resolution for detailed species classification and biomass estimation at the field scale [16,17,18,19]. Studies reveal strong correlations between vegetation indices derived from PlanetScope and UAV data, underscoring their synergistic potential for multi-scale ecological assessments [1,20]. In this regard, integrating these platforms that take advantage of UAVs’ precise spatial precision and satellite data’s broad coverage is required to build robust and scalable grassland monitoring and management systems.

In the past decade, deep learning-based super-resolution (SR) methods have created new opportunities for downscaling remote sensing imagery from low to higher spatial resolution [21,22,23,24,25]. Early convolutional neural network (CNN) approaches demonstrated strong capability for learning spatial features [26,27], but more recent methods based on Generative Adversarial Networks (GANs) have substantially improved the reconstruction of fine-scale textures. GANs employ an adversarial training scheme in which a generator produces high-resolution images from low-resolution inputs while a discriminator evaluates their realism, encouraging the synthesis of perceptually plausible details. Super-resolution architectures such as SRGAN and its enhanced variant ESRGAN exploit this framework to learn complex non-linear mappings and generate realistic textures, often outperforming interpolation and conventional CNN-based models in preserving fine ecological features [28,29,30]. GAN-based approaches have been widely applied in domains including medical image enhancement, satellite image analysis, surveillance, and infrastructure inspection [31]. More recently, vegetation-oriented models such as VegGAN [32] and extreme geospatial downscaling approaches [33] reflect growing interest in ecological applications. However, while SR methods have been extensively applied to structured landscapes, their performance in heterogeneous natural environments such as grasslands remains comparatively underexplored.

While GANs have emerged as powerful tools for downscaling remote sensing imagery, a critical gap remains in understanding how intrinsic landscape properties—spatial heterogeneity and textural complexity—influence their performance in natural ecosystems. Most evaluations rely on global metrics averaged across entire datasets, overlooking how algorithm performance varies under different landscape conditions [23,34]. This is problematic for grasslands, where ecological processes operate across multiple scales and management requires reliable fine-grain information [35,36]. The assumption that GANs perform uniformly across landscapes remains untested; theoretical and empirical evidence suggests that spatial structure, vegetation density, and textural complexity strongly influence the information content available for resolution enhancement [3,29,37,38]. Although textures are critical for SR, no study has quantified how fine-scale textural complexity—reflecting vegetation variability—affects SR fidelity [23,39].

In parallel, the robustness of GAN-based downscaling under different operational configurations—hereafter referred to as Downscaling Strategy Frameworks—remains insufficiently examined. Key scenarios, including intra-sensor downscaling, cross-sensor generalization, and domain adaptation, introduce substantial distribution shifts that challenge model transferability and often lead to pronounced performance degradation [40,41]. These scenarios are central to real-world applications, where training and deployment conditions rarely align; however, the absence of integrated evaluations that jointly account for landscape heterogeneity and operational strategy constrains both model interpretability and practical deployment in heterogeneous natural environments.

Our research addresses these gaps by comparing SRGAN [30] and ESRGAN [28] against bicubic interpolation. We evaluate performance using PSNR and SSIM across stratified NDVI, structure, and texture gradients, providing the first landscape-aware assessment of GAN-based SR under Downscaling Strategy Frameworks. NDVI captures spectral gradients critical for image reconstruction [2]; spatial structure, quantified through patch metrics, reflects habitat configuration and biodiversity patterns [3,11]; and textural complexity, measured using GLCM entropy, captures fine-scale variability beyond spectral means and relates to plant richness [38]. By integrating these axes into a stratification framework, we move beyond global accuracy metrics toward ecologically informed evaluation. To our knowledge, this is the first study applying such a stratified approach to GAN-based SR in natural ecosystems, grounding methodological innovation in ecological theory and enhancing grassland monitoring applications.

The objective of this study is to evaluate how landscape heterogeneity influences the performance and robustness of super-resolution models in grassland ecosystems under realistic operational conditions. Rather than conducting a broad benchmarking of competing architectures, we adopt a controlled experimental design to isolate the effects of ecological complexity and domain shift on reconstruction fidelity. Specifically, we examine how gradients in vegetation condition, spatial structure, and textural complexity affect model performance across three downscaling strategy frameworks: intra-sensor, cross-sensor, and intra-to-cross generalization. Two representative GAN-based models, SRGAN and ESRGAN, are selected as canonical architectures with comparable foundations but differing capacities for high-frequency texture reconstruction, enabling targeted assessment of whether increased model sophistication improves resilience to heterogeneity. A bicubic interpolation baseline is included to contextualize deep learning performance under varying conditions. By explicitly linking landscape characteristics with sensor-transfer scenarios, this study advances a landscape-aware evaluation framework that identifies when and why super-resolution approaches succeed or fail in heterogeneous grassland environments. In doing so, it shifts the focus from model-centric comparison toward understanding the environmental constraints that govern the reliability of super-resolution for ecological applications.

2. Methodology

To investigate how landscape heterogeneity and domain shift influence super-resolution performance, we adopt a controlled experimental framework that isolates environmental and operational factors from architectural variability. Rather than benchmarking multiple models, the focus is on how reconstruction behavior changes across ecological conditions and deployment scenarios. We therefore select two representative GAN architectures—SRGAN and ESRGAN—that share a common formulation but differ in their ability to reconstruct high-frequency textures, enabling a targeted assessment of robustness to landscape complexity. A bicubic interpolation approach is included as a baseline. Model performance is evaluated across stratified gradients of vegetation condition, spatial structure, and textural complexity, and under three downscaling strategy frameworks (intra-sensor, cross-sensor, and intra-to-cross generalization), allowing us to quantify how heterogeneity and domain mismatch jointly influence super-resolution fidelity in grassland systems.

2.1. Study Area

The study area is located within the Edwards Plateau ecoregion at the Carl and Bina Sue Martin—Texas A&M AgriLife Research Ranch (MR), Menard, Texas (30.8096°N, −99.8657°W), covering approximately 1553 ha (Figure 1). The landscape is a Mesquite-Oak savanna, with woody species such as honey mesquite, live oak, agarita, and Ashe juniper. Common grasses include sideoats grama, Texas wintergrass, and Aristida species. Soils are primarily of the Tarrant series, characterized by very cobbly silty clay and clay with 1–15% slopes. The region has a humid subtropical climate, with temperatures ranging from 7.9 °C (January) to 34.4 °C (August) and annual rainfall averaging 640 mm, peaking in June. Mixed herds of goats, cattle, and sheep graze year-round under a pyric herbivory regime. Prescribed burns are conducted annually over about 200 ha across two rotating burn units.

2.2. Overview of the Methodological Approach

Figure 2 provides an integrated overview of the methodological workflow implemented in this study, summarizing the end-to-end data pipeline from image acquisition to model evaluation. The workflow begins with data collection, followed by preprocessing and tile extraction. These paired tiles serve as the input to the downscaling models, which are trained under multiple experimental configurations. Model performance is subsequently evaluated using standard image quality metrics, including Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index (SSIM). The final outputs of the modeling phase are downscaled images with enhanced spatial detail.

To rigorously assess how downscaling performance varies across landscape heterogeneity, the study employs a structured set of stratification scenarios. First, vegetation condition was examined by grouping images into quartiles based on NDVI. Second, landscape structure was evaluated by clustering images according to patch-based metrics—total number of woody patches, mean patch area, and the standard deviation of patch area. Third, textural complexity was assessed by classifying images into quartiles based on entropy. Beyond stratification, three downscaling strategy frameworks were tested: intra-sensor downscaling, cross-sensor downscaling, and intra-to-cross generalization. Finally, three models were compared—SRGAN, ESRGAN, and Bicubic interpolation. The multiple experimental configurations are shown in Figure 3 and described in detail in the following subsections.

2.3. Data Sources and Preprocessing

2.3.1. UAV and Satellite Imagery Acquisition

High-resolution imagery was collected using an unmanned aerial vehicle (UAV) platform as part of a structured remote sensing campaign designed to support spatial downscaling analysis. UAV flights were conducted between April and June 2025 along predefined, overlapping flight paths—each approximately 100 m wide and varying in length from 800 to 5000 m. The flight paths also differed in form, including linear and zigzag patterns, to capture landscape heterogeneity and ensure dense spatial coverage (Figure 4b). Flight missions were planned and executed autonomously using WingtraHub 1.0 software (Wingtra, Zurich, Switzerland). Image acquisition was performed using a WingtraOne Gen II fixed-wing UAV equipped with a MicaSense RedEdge-P multispectral sensor (AgEagle Aerial Systems, Seattle, WA, USA). The sensor captured five spectral bands: blue (B), green (G), red (R), red-edge (RE), and near-infrared (NIR). Flights were conducted at an altitude of 60 m with 75% forward and side overlap, ensuring sufficient image for accurate orthomosaic generation and reflectance consistency. The resulting UAV orthomosaics achieved a ground sampling resolution of approximately 0.03 × 0.03 m.

In parallel, PlanetScope imagery was acquired to serve as the low-resolution counterpart in the downscaling framework (Figure 4a). PlanetScope data, with a spatial resolution of 3 × 3 m, offers near-daily revisit capabilities and includes bands compatible with those captured by the UAV sensor. To reduce phenological and illumination mismatches between datasets, efforts were made to align acquisition dates as closely as possible between the UAV flights and satellite imagery.

2.3.2. Preprocessing and Tiling Strategy

Initial preprocessing of the UAV imagery involved radiometric calibration and orthomosaic generation using manufacturer-recommended procedures. Particular attention was given to geometric accuracy and consistent reflectance values across tiles to ensure high-quality reference data. These calibrated orthomosaics formed the high-resolution ground truth dataset used for training and evaluation.

A GIS-based approach was adopted to generate paired input-output datasets for model training. Square sampling polygons measuring 60 × 60 m were randomly distributed across the UAV imagery, with some degree of overlap allowed to ensure comprehensive spatial representation (Figure 5a). For each polygon, corresponding tiles were extracted from both the UAV (high-resolution) and PlanetScope (low-resolution) imagery using the Clip function in ArcGIS Pro (version 3.4), automated via Python scripting (version 3.9.23) for batch processing (Figure 5b). Each tile pair was assigned a unique identifier for traceability and model input management. As a result, a total of 5142 pairs of tiles were obtained.

To ensure compatibility and reduce model complexity, only the RGB bands were retained from both datasets. The resulting dataset of paired image tiles served as the input (low resolution) and ground truth (high resolution) for the deep learning-based super-resolution modeling workflow.

2.4. Landscape Stratification: Tile-Based Classification for Downscaling Models

To evaluate the influence of landscape characteristics on downscaling model performance, tile images were stratified into discrete classes based on vegetation health, landscape structure, and texture. These classifications enabled a comparative assessment of model behavior across gradients of ecological and spatial complexity. The rationale behind this stratification is grounded in the expectation that spatial heterogeneity, vegetation density, and textural complexity may influence the capacity of super-resolution models to reconstruct fine-scale patterns. All stratification layers (NDVI, structure, and texture) were derived from the same set of image tiles, ensuring that comparisons across dimensions are directly comparable and not affected by differences in sampling.

2.4.1. Vegetation Health (NDVI Quartile)

Vegetation health condition was quantified using the Normalized Difference Vegetation Index (NDVI), a widely used proxy for photosynthetic activity and plant vigor. NDVI values were computed for each of the 5142 tile images, and the distribution was divided into four quartiles: Q1 (lowest NDVI), Q2, Q3, and Q4 (highest NDVI). This quartile-based classification enabled separate evaluations of model performance across vegetation conditions, from sparsely vegetated to densely vegetated regions. Since NDVI is correlated with both canopy condition and spectral variability, the quartile approach allowed us to assess whether models perform differently in areas with low versus high biomass. Because quartiles inherently produce near-equal sample sizes, no additional resampling was required, and the standard 80/20 training–testing split was applied within each quartile. Quartile statistics (mean and standard deviation) are summarized in

Table 1. Statistics and landscape characterization of NDVI-based quartile classification of image tiles. Quartiles were defined based on mean NDVI values computed per tile. The classified tiles reflect a gradient from bare or sparsely vegetated ground (Q1) to areas with dense vegetation and tree canopy (Q4).

Quartile	Pattern	Tile Count	NDVI Mean (SD)	Characterization
Q1	Very Low Vegetation Cover	1286	0.25 (±0.022)	Dominated by bare soils or sparsely vegetated areas; low photosynthetic activity.
Q2	Low to Moderate Cover	1285	0.31 (±0.014)	Likely areas with patchy grassland or early-stage regrowth; slightly greener zones.
Q3	Moderate Vegetation Cover	1285	0.36 (±0.022)	Predominantly grassland with some canopy presence; increased uniformity.
Q4	Dense Vegetation Canopy	1286	0.47 (±0.052)	Likely dominated by tree-covered patches or mixed grassland with high biomass.

, and representative tiles illustrating vegetation differences across quartiles are shown in Figure 6.

2.4.2. Landscape Structure (Clustering Using Patch Metrics)

Landscape structure was characterized based on woody vegetation patterns using three patch-based metrics: total number of woody patches, mean patch area, and standard deviation of patch area. To extract these metrics, the 5142 UAV tile images were converted to binary vegetation masks using an NDVI threshold of 0.50 and a minimum patch size of 200 pixels. These thresholds ensured consistent delineation of woody features across tiles.

The resulting structural descriptors captured landscape fragmentation and aggregation, which influenced spatial texture and signal complexity in remote sensing imagery. A k-means clustering algorithm was applied to the three patch-based metrics, and the optimal number of clusters (k = 3) was determined using the elbow method (Details of this analysis are included as Supplementary Material). The elbow curve (Figure S1) shows an inflection point between k = 3 and k = 4; k = 3 was selected to provide a parsimonious and interpretable representation of landscape structural variability.

The identified clusters were characterized as follows:

• Cluster 1 (C1)—Sparse and Fragmented Woody Patches: Characterized by low woody cover, with numerous small, isolated patches embedded in herbaceous or bare-ground matrices. This was the most common class in the study area.
• Cluster 2 (C2)—Dispersed Woody Mosaics: Comprised of many small to medium patches forming a heterogeneous mosaic. These areas are indicative of ecotones or transitional savanna states.
• Cluster 3 (C3)—Dense and Clumped Woody Dominance: Represented by large, contiguous woody patches with high canopy closure, indicative of mature stands or late-stage woody encroachment.

Cluster statistics are reported in

Table 2. Statistics and landscape features of image tiles grouped by structure-based clustering using patch-level metrics: tree patch area, number of patches, and patch area variability.

Cluster	Name	Tile Count	Tree Patch Count Mean (SD)	Mean Object Size (Pixels2) Mean (SD)	Patch Size Std Dev (Pixels2) Mean (SD)	Characterization
C1	Sparse & Fragmented Woody Patches	3179	22.7 (±10.9)	5212 (±3975)	11,794 (±11,445)	Few, small, scattered tree patches in bare/herbaceous matrix; high fragmentation; early encroachment or naturally sparse woody growth.
C2	Dispersed Woody Mosaics	1452	58.2 (±16.9)	4655 (±2202)	15,022 (±12,364)	Very high number of small woody patches forming a mosaic; abundant but not dense; advanced but patchy encroachment or savanna structure.
C3	Dense & Clumped Woody Dominance	511	21.1 (±10.2)	24,691 (±18,294)	78,123 (±40,569)	Large, contiguous woody patches; low count but very large and dominant; dense forest/shrubland or late-stage encroachment.

, and representative tile examples for each cluster are shown in Figure 7.

Because the three clusters contained different numbers of tiles, we selected equal numbers of training and testing samples from each group to ensure a balanced evaluation. This stratification produced a balanced training set of 409 paired tiles (80%) and a test set of 102 paired tiles per cluster (20%), enabling consistent comparisons of reconstruction accuracy across landscape-structure classes. This balancing step was applied only to the structure-based stratification to mitigate class imbalance, whereas NDVI and entropy stratifications retained their full sample sizes due to their balanced quartile design.

2.4.3. Texture (Entropy-Based Stratification)

Textural complexity was quantified using Gray-Level Co-occurrence Matrix (GLCM) Entropy on UAV tile images (high resolution imagery; cell size = 0.03 × 0.03 m). A 33-pixel displacement (~1 m) in four directions (0°, 45°, 90°, 135°) was used to capture the spatial arrangement of vegetation elements (tree crowns, shrub clusters, and grass patches) while minimizing fine-scale pixel noise. This scale approximates the typical canopy size in mesquite–oak savanna, consistent with standard GLCM approaches for local structural variability [42]. Entropy values were divided into quartiles (Q1–Q4), yielding approximately equal sample sizes, and the same 80/20 training–testing split was applied within each quartile without additional resampling. Quartile statistics are summarized in

Table 3. Summary statistics and landscape descriptions for the texture-based quartile classification of image tiles. Quartiles are derived from mean entropy values per tile, capturing a gradient from low textural disorder and simpler vegetation patterns (Q1) to highly complex and spatially disordered landscapes (Q4).

Quartile	Pattern	Tile Count	NDVI Mean (SD)	Characterization
Q1	Lowest Spatial Disorder (Very Low Textural Complexity)	1286	11.82 (±0.47)	Large, homogeneous tonal areas; minimal fine-scale variation. Represents open grasslands or sparsely vegetated surfaces with few woody elements.
Q2	Moderate Tonal Variability (Low-to-Moderate Textural Complexity)	1285	12.83 (±0.20)	More frequent tonal changes, but still relatively simple patterns. Transitional zones with scattered shrubs or small tree patches beginning to break the grassland matrix.
Q3	High Pattern Complexity (Moderate-to-High Textural Complexity)	1285	13.50 (±0.17)	Strong tonal intermixing; frequent contrast transitions. Heterogeneous vegetation mosaics with fragmented tree canopies and mixed grass–shrub structure.
Q4	Maximum Spatial Disorder (Highest Textural Complexity)	1286	13.99 (±0.16)	Highly disordered tonal patterns with dense fine-scale variation. Densely vegetated areas with complex canopy layering, abundant shadows, and highly diverse plant structure.

, and representative tiles illustrating texture differences across quartiles are shown in Figure 8.

2.5. Downscaling Strategy Frameworks

To evaluate the performance and generalization capability of the downscaling models, three distinct downscaling strategy frameworks were defined: intra-sensor downscaling, cross-sensor downscaling, and intra-to-cross generalization. The three downscaling strategy frameworks are illustrated schematically in Figure 9 to clarify differences in training and testing configurations. These scenarios reflect different operational contexts and data availability conditions in remote sensing applications.

2.5.1. Intra-Sensor Downscaling (UAV to LR + HR)

The intra-sensor strategy involves using high-resolution (HR) and synthetically degraded low-resolution (LR) imagery from the same sensor platform. In this setup, tile images collected by the UAV were artificially downsampled by a factor of four (×4) to generate LR versions. These LR-HR image pairs were then used to train and evaluate the downscaling models. This scenario allows for controlled testing and minimizes sensor-related spectral or spatial inconsistencies, providing a benchmark to assess model fidelity under ideal conditions.

2.5.2. Cross-Sensor Downscaling (Planet as LR, UAV as HR)

The cross-sensor strategy explores a more practical application where the low-resolution images are obtained from the PlanetScope satellite platform, while the high-resolution reference images come from UAV imagery. In this case, HR and LR imagery originate from different sensors, each with unique spatial, spectral, and radiometric characteristics. This setup introduces additional complexity due to inter-sensor differences, providing insight into how well the models perform in real-world multi-sensor applications.

2.5.3. Intra-to-Cross- Generalization (Trained on UAV, Applied to Planet)

In the intra-to-cross generalization strategy, the models were initially trained using the intra-sensor UAV-derived LR-HR tile pairs (as in Section 2.5.1) and then applied to downscale PlanetScope LR images. The resulting downscaled images were evaluated against co-located UAV HR images. This approach examines the potential of model generalization across sensor domains, assessing whether models trained on only UAV imagery can effectively enhance satellite images. The scenario is particularly relevant for operational workflows where HR ground truth data is limited and model re-training on new sensors is not feasible.

2.6. Downscaling Models

To assess how vegetation health, landscape-structure heterogeneity, and textural complexity within grassland mosaics influence the performance of image downscaling, this study compares three methods for enhancing spatial resolution: SRGAN, ESRGAN, and bicubic interpolation. Bicubic interpolation, a widely used conventional technique, serves as a baseline, while SRGAN and ESRGAN are deep learning-based super-resolution models designed to reconstruct fine-scale details beyond the capabilities of traditional interpolation. All models were applied to RGB image tiles using a consistent ×4 upscaling factor.

2.6.1. Super-Resolution Generative Adversarial Network (SRGAN) Model

The SRGAN, proposed by [30], is a deep learning framework designed to generate high-resolution (HR) images from low-resolution (LR) inputs while producing perceptually realistic textures. SRGAN follows the Generative Adversarial Network (GAN) paradigm [43], in which two neural networks (a generator and a discriminator) are trained adversarially: the generator produces super-resolved images, while the discriminator learns to distinguish generated images from real HR examples (Figure 10).

The generator $G$ learns a mapping from LR images $I_{LR}$ to super-resolved outputs $I_{SR}$, such that $I_{SR}=G\left(I_{LR}\right)$ and $I_{SR}\approx I_{HR}$. Architecturally, the generator is based on a deep residual network (SRResNet) consisting of an initial convolutional layer followed by multiple residual blocks. Each block includes two convolutional layers, batch normalization, and Parametric ReLU (PReLU) activations with skip connections to facilitate gradient flow. After the residual blocks, sub-pixel convolution (PixelShuffle) layers progressively increase spatial resolution (in two stages for a ×4 upscaling) followed by a final convolutional layer producing the super-resolved image.

The discriminator D serves as a binary classifier trained to differentiate between real HR images $I_{HR}$ and generated SR images $I_{SR}$. It consists of a series of convolutional layers with increasing feature depth, interleaved with batch normalization and Leaky ReLU activations, progressively reducing spatial dimensions. The final layers are fully connected, culminating in a sigmoid output indicating the probability that the input image is real. The discriminator’s architecture resembles that of the VGG family of networks, with small receptive fields (e.g., 3 × 3) and hierarchical feature extraction.

$$D\left(I\right)=\left\{\begin{array}{l}1, if I is a real HR image\\ 0, if I is a generated SR image\end{array}\right.$$

Training SRGAN involves optimizing a perceptual loss, which combines adversarial loss and content loss. The adversarial loss encourages the generator to produce images that reside on the manifold of natural images, as judged by the discriminator, and is defined as:

$${\mathcal{L}}_{adv} =-\log D\left(G\left(I_{LR}\right)\right),$$

where $D\left(G\left(I_{LR}\right)\right)$ is the discriminator’s estimate of the probability that $G\left(I_{LR}\right)$ is a real image.

The content loss—computed as the Euclidean distance between feature representations of $I_{SR}$ and $I_{HR}$ extracted from a pre-trained VGG network—ensures that the generated images preserve perceptually important structures/content from the HR image.

$${\mathcal{L}}_{conmtent}={\displaystyle \frac{1}{WHC}} {‖\varphi \left(I_{SR}\right)-\varphi \left(I_{HR}\right)‖}_2^2,$$

where $\varphi \left(\cdot \right)$ denotes the feature maps extracted from a pre-trained VGG19 network, and W, H, and C are the spatial dimensions and channel count of the feature maps. The total perceptual loss used to train the generator is:

$${\mathcal{L}}_G={\mathcal{L}}_{content}+\lambda \hspace{0.17em}{\mathcal{L}}_{adv},$$

where λ balances perceptual fidelity and image realism.

For this study, SRGAN was trained using co-registered PlanetScope (LR) and UAV (HR) image tiles. The trained model was then applied to PlanetScope imagery to generate ×4 super-resolved outputs, which were evaluated against UAV imagery to assess spatial accuracy and the ability to recover ecologically relevant detail [44].

2.6.2. Enhanced Super-Resolution Generative Adversarial Network (ESRGAN) Model

The Enhanced Super-Resolution Generative Adversarial Network (ESRGAN), proposed by Wang et al. [28], replaces SRGAN’s residual blocks with Residual-in-Residual Dense Blocks (RRDB), which integrate dense connections and multi-level residual learning without batch normalization. This deeper and more stable generator enables the reconstruction of high-frequency details, which is an important advantage for remote sensing applications requiring fine spatial structure.

ESRGAN also introduces a relativistic discriminator, which evaluates whether real images appear more realistic than generated ones, rather than classifying images independently. This formulation stabilizes adversarial training and encourages more naturalistic textures. In addition, ESRGAN improves the perceptual loss using VGG feature maps pre-activation, enhancing structural fidelity relative to SRGAN.

In this study, ESRGAN was trained using the same co-registered PlanetScope–UAV patch pairs as SRGAN, enabling direct comparison across downscaling framework strategies. The super-resolved outputs were evaluated against UAV imagery to assess their performance.

Both SRGAN and ESRGAN models were trained using the Adam optimizer. The generator was optimized with a learning rate of 1 × 10⁻⁴ (β₁ = 0.5), while the discriminator used a learning rate of 4 × 10⁻⁴ (β₁ = 0.9), both with gradient clipping (clipnorm = 0.1). Training was conducted for 30 epochs with a batch size of 1. A ReduceLROnPlateau scheduler was applied to the generator, with a decay factor of 0.1, patience of 5 epochs, and a minimum learning rate of 1 × 10⁻⁶. The combined generator loss was formulated as $L={\lambda }_{adv}\cdot L_{adv}+{\lambda }_{vgg}\cdot L_{perceptual}+{\lambda }_{pixel}\cdot L_{pixel}+{\lambda }_{tv}\cdot L_{tv}$, where the weights were set as: ${\lambda }_{adv}$ = 1 × 10⁻³, ${\lambda }_{vgg}$ = 0.5, ${\lambda }_{pixel}$ = 1.0, and ${\lambda }_{tv}$ = 1 × 10⁻⁶. The perceptual loss was computed using features extracted from layer 10 of a VGG19 network pre-trained on ImageNet. Table S1 in the Supplementary Material shows list all the hyperparameters.

2.6.3. Bicubic Interpolation

Bicubic interpolation is a classical image resampling technique based on cubic convolution, commonly used as a baseline in image super-resolution tasks [45]. It estimates the intensity of a new pixel by fitting a bicubic surface to the 16 nearest pixel values in a 4 × 4 neighborhood. This method assumes the image surface is continuous and differentiable, allowing smoother transitions in upscaled images. The interpolated pixel value $I\left(x,y\right)$ at a continuous spatial location $\left(x,y\right)$ is computed as:

$$I\left(x,y\right)=\sum _{i=-1}^2\sum _{j=-1}^{2}w\left(i,j\right)\cdot I\left(x_i,y_j\right),$$

where $I\left(x_i,y_j\right)$ are the known neighboring pixel values and $\left(i,j\right)$ are the bicubic interpolation weights, typically derived from cubic convolution kernels), given by:

$$w\left(x\right)=\left\{\begin{array}{c}\begin{array}{l}\left(a + 2\right){\left|x\right|}^3- \left(a + 3\right){\left|x\right|}^2+ 1, \mathrm{i}\mathrm{f} \left|x\right|< 1\\ a {\left|x\right|}^3- 5a {\left|x\right|}^2+ 8a\left|x\right|- 4a, \mathrm{i}\mathrm{f} 1 \le \left|x\right|< 2\end{array}\\ 0, \mathrm{otherwise}\end{array}\right.$$

(1)

with $a=-0.5$ being a commonly used parameter (Catmull–Rom spline). In this study, bicubic interpolation is used as a non-learning-based benchmark to assess the performance of learning-based super-resolution models, particularly SRGAN and ESRGAN. Unlike these deep learning models, bicubic interpolation does not involve training and does not adapt to image content, providing a fixed reference for evaluating image reconstruction quality through metrics such as PSNR, SSIM, MAE, and MSE.

2.7. Model Performance Measurement (Evaluation Metrics)

To evaluate the performance of the super-resolution model, we employed three widely recognized metrics: Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index (SSIM).

Peak Signal-to-Noise Ratio (PSNR) is a metric that evaluates the ratio between the maximum possible signal power and the noise introduced by compression or prediction errors. A higher PSNR generally corresponds to better image quality. Commonly used thresholds for PSNR quality are as follows: 20 dB indicates poor quality, 30 dB represents acceptable quality, and values above 40 dB indicates high-quality images. PSNR is given by the following formula:

$$\mathrm{PSNR}=10\cdot {\log }_{10}\left({\displaystyle \frac{{\mathrm{MAX}}^2}{\mathrm{MSE}}}\right),$$

where MAX is the maximum possible pixel value (for example, 255 for 8-bit images), and MSE is as defined above.

The Structural Similarity Index (SSIM) assesses the perceptual similarity between two images, accounting for luminance, contrast, and structural patterns. SSIM values range from −1 to 1, with a value of 1 indicating perfect structural similarity. The SSIM formula is defined as:

$$\mathrm{SSIM}\left(x,\widehat{x}\right)={\displaystyle \frac{\left(2{\mu }_x{\mu }_{\widehat{x}}+C_1\right)\left(2{\sigma }_{x\widehat{x}}+C_2\right)}{\left({\mu }_x^2+{\mu }_{\widehat{x}}^2+C_1\right)\left({\sigma }_x^2+{\sigma }_{\widehat{x}}^2+C_2\right)}},$$

where ${\mu }_x$ and ${\mu }_{\widehat{x}}$ represents the means of the images, ${\sigma }_x^2$ and ${\sigma }_{\widehat{x}}^2$ denote the variances, ${\sigma }_{x\widehat{x}}$ is the covariance, and $C_1$ and $C_2$ are constants added to stabilize the computation.

Quantifying levels of variation (Coefficient of variation). To assess the internal variability of image reconstruction performance within each quartile (for vegetation-health and entropy-based segmentation) or within each cluster (for entropy-cluster segmentation), we used the coefficient of variation (CV), defined as the ratio of the standard deviation to the mean (CV = SD/M). In this study, CV < 10% was classified as low variability and high consistency, 10% < CV < 20% as moderate variability, and CV ≥ 20% as high variability or instability. CV values above ~30% were interpreted as indicating substantial dispersion and reduced reliability in model performance.

Perceptual metrics such as Fréchet Inception Distance (FID) were not included, as they rely on feature representations learned from natural image datasets and require large sample sizes for stable estimation [46]. These conditions are not fully met in multispectral remote sensing contexts, limiting their interpretability

3. Results

The results are organized according to the three landscape stratification schemes examined in this study: vegetation health (NDVI-based), landscape structure, and landscape texture. Within each scheme, downscaling performance is evaluated across the three downscaling frameworks (intra-sensor, cross-sensor, and intra-to-cross generalization). The following subsections present the results for each stratification, highlighting how vegetation condition, structural configuration, and texture complexity influence model performance.

To evaluate potential redundancy among stratification variables, we examined their relationships. NDVI and entropy showed a moderate-to-high correlation (r = 0.79), indicating partial overlap between spectral condition and textural variability. However, entropy differed significantly across NDVI quartiles (Kruskal–Wallis, p < 0.001), increasing from 9.67 in Q1 to 11.18 in Q4, indicating that textural complexity is not solely determined by vegetation greenness. Correlations between these variables and patch-based structural metrics were lower (r = 0.30–0.68). In addition, NDVI differed significantly across structure clusters (Kruskal–Wallis, p < 0.001), with mean NDVI values of 0.304, 0.403, and 0.465 for clusters C1, C2, and C3, respectively. These results indicate that the stratification variables are related but not redundant, capturing complementary aspects of landscape heterogeneity.

Overall, downscaling accuracy varied across stratification schemes, downscaling frameworks, and algorithms. The following subsections present these results in detail, highlighting the influence of vegetation condition, structural configuration, and texture complexity on model performance.

3.1. Downscaling Performance Across NDVI-Derived Quartiles

The results in this subsection evaluate the performance of the proposed downscaling strategies using the independent testing dataset. As described in the methodology, the dataset was stratified into four quartiles (Q1–Q4), comprising 1286, 1285, 1285, and 1286 tile images, respectively. From these, 258, 257, 257, and 258 images were allocated to the testing subset of Q1–Q4, ensuring balanced evaluation across quartiles. Accordingly, the results presented here correspond to the reconstruction performance of the respective number of testing images in each quartile.

3.1.1. Cross-Quartile Downscaling Performance Based on NDVI

The evaluation of cross-quartile downscaling revealed distinct performance patterns across the examined models and strategy frameworks, indicating a clear influence of vegetation health on the downscaling process. Figure 11 illustrates comparative boxplots of peak signal-to-noise ratio (PSNR) values across the evaluated downscaling algorithms, downscaling strategy frameworks, and NDVI-based quartiles.

Table 4. Mean, standard deviation, and coefficient of variation in PSNR and SSIM values across NDVI-based quartiles, reported for each algorithm (SRGAN, ESRGAN, and Bicubic) and by downscaling strategy framework (intra-sensor, cross-sensor, and intra-to-cross generalization).

	SR GAN Model												ESR GAN Model												BICUBIC Model
Metric	Intra-Sensor				Cross-Sensor				Generalization				Intra-Sensor				Cross-Sensor				Generalization				Intra-Sensor				Cross-Sensor
	Q1	Q2	Q3	Q4	Q1	Q2	Q3	Q4	Q1	Q2	Q3	Q4	Q1	Q2	Q3	Q4	Q1	Q2	Q3	Q4	Q1	Q2	Q3	Q4	Q1	Q2	Q3	Q4	Q1	Q2	Q3	Q4
PSNR Mean	34.74	33.64	27.45	27.92	28.43	28.30	27.86	25.94	21.07	4.45	9.41	3.26	34.93	35.26	35.26	36.46	29.02	28.97	29.01	30.07	17.39	18.06	20.21	20.78	36.56	36.58	36.62	37.94	17.52	20.08	20.01	19.95
PSNR SD	1.92	6.26	10.35	12.93	2.14	2.12	2.01	2.99	2.85	4.53	6.29	3.64	2.13	1.80	1.81	2.42	2.16	1.82	1.57	1.95	2.47	3.23	3.44	3.23	2.50	2.14	2.04	2.54	2.82	2.96	3.16	3.45
PSNR CV	5.52	18.56	37.62	46.22	7.50	7.49	7.21	11.50	13.54	101.8	66.81	111.4	6.10	5.10	5.14	6.63	7.45	6.28	5.41	6.47	14.19	17.86	17.01	15.55	6.85	5.86	5.58	6.70	16.11	14.76	15.79	17.29
SSIM Mean	0.83	0.84	0.81	0.81	0.77	0.77	0.76	0.71	0.71	0.29	0.43	0.27	0.82	0.83	0.84	0.86	0.76	0.74	0.76	0.79	0.51	0.59	0.61	0.58	0.86	0.87	0.87	0.89	0.60	0.63	0.62	0.64
SSIM SD	0.06	0.06	0.10	0.15	0.06	0.05	0.05	0.08	0.07	0.15	0.16	0.14	0.07	0.05	0.05	0.05	0.06	0.05	0.04	0.06	0.07	0.09	0.09	0.11	0.06	0.05	0.04	0.04	0.08	0.08	0.07	0.09
SSIM CV	7.74	7.61	11.84	18.68	7.61	6.53	5.90	10.87	9.80	49.69	36.79	52.45	8.35	6.34	5.65	6.21	8.46	6.96	5.33	7.44	14.55	15.07	14.72	18.23	7.49	5.85	4.94	4.80	13.46	12.47	11.97	14.10

shows the mean, standard deviation, and coefficient of variation (CV) of PSNR and SSIM values across NDVI-based quartiles. Figure S3 of the Supplementary Material illustrates comparative boxplots of SSIM values.

For the SRGAN model under the intra-sensor framework, Quartiles 1 and 2, corresponding to low NDVI values, achieved mean PSNR levels above 33 dB with relatively narrow distributions, indicating stable reconstructions in sparsely vegetated areas. In contrast, Quartiles 3 and 4, associated with moderate to high NDVI, exhibited substantial variability, with wide interquartile ranges, including PSNR values below 15 dB. These results suggest that SRGAN struggles to maintain reconstruction quality in vegetation-dense regions, where higher NDVI reflects greater canopy cover and structural complexity, including overlapping crowns and shadowing effects that increase spatial heterogeneity in the imagery. Normality assumptions were not satisfied, and the Kruskal–Wallis’ test confirmed significant differences in performance across quartiles (H(3) = 38.13, p < 0.001). For the SRGAN model under the cross-sensor framework, performance across quartiles also differed significantly (H(3) = 132.57, p < 0.001). However, the decrease in PSNR with increasing NDVI was less pronounced compared to the intra-sensor case. Variability remained relatively consistent across Quartiles 1–3 (CV ≈ 7.5%), with a modest increase in Quartile 4 (CV = 11.5%), indicating some instability in vegetation-dense regions but an overall more controlled performance across NDVI classes. For the SRGAN model under the intro-to-cross generalization framework, the model performance deteriorated dramatically. Quartile 1 achieved a mean PSNR of only 21.07 dB, while Quartiles 2–4 recorded severely degraded averages of 4.45, 9.41, and 3.26 dB, respectively. These outcomes were accompanied by extreme variability, with coefficients of variation of 13.5%, 101.8%, 66.8%, and 111.4% for Quartiles 1–4, respectively. The results indicate that SRGAN fails to generalize effectively across sensors when applied to NDVI-derived classes. Kruskal–Wallis testing confirmed significant differences among quartiles in this scenario as well (H(3) = 651.04, p < 0.001). Overall, these findings demonstrate that SRGAN model performance is strongly affected by the NDVI-based vegetation density of the scene, with higher NDVI regions posing the greatest challenges and exhibiting the largest variability across all scenarios.

For the ESRGAN model evaluated under the intra-sensor framework, downscaling performance showed a slight increasing trend from low to high NDVI conditions. Quartile 4 produced marginally higher average PSNR values (36.46 dB) compared to Quartile 3 (35.26 dB), while Quartiles 2 and 3 were virtually indistinguishable (both 35.26 dB), and Quartile 1 yielded the lowest average (34.93 dB). Although the mean and distributional ranges of PSNR values appeared visually similar across some quartiles, the Kruskal–Wallis test confirmed statistically significant differences among them (H(3) = 62.85, p < 0.001). Under the cross-sensor framework, ESRGAN performance across NDVI-derived quartiles was more uniform, with average PSNR values ranging narrowly from 28.97 to 30.07 dB, and the highest values observed in Quartile 4. Despite the small numerical separation among quartiles, the Kruskal–Wallis test again indicated significant differences in performance (H(3) = 69.32, p < 0.001). For the intra-to-cross generalization framework, ESRGAN showed a clear monotonic improvement from Q1 to Q4. Average PSNR values increased from 17.39 dB (Q1) to 18.06 dB (Q2), 20.21 dB (Q3), and 20.78 dB (Q4). These differences were strongly significant according to the Kruskal–Wallis test (H(3) = 206.04, p < 0.001). Across all three evaluation frameworks, variability in ESRGAN performance remained relatively low: coefficients of variation ranged from 5.1–6.6% in the intra-sensor framework, 5.4–7.4% in the cross-sensor framework, and 14.2–17.9% in the intra-to-cross generalization framework. Overall, the results indicate that ESRGAN performance is generally stable across NDVI-derived vegetation classes and shows a gradual improvement as NDVI increases. Higher NDVI values correspond to denser vegetation cover and greater canopy continuity, which tends to produce more homogeneous spectral responses while maintaining spatial structure associated with vegetation canopies. As a result, reconstruction accuracy improves slightly under greener canopy conditions across all evaluation frameworks, suggesting that ESRGAN is better able to preserve structured vegetation patterns even under dense canopy conditions, highlighting ESRGAN’s robustness in grassland-dominated environments with varying levels of vegetation vigor.

The Bicubic model under the intra-sensor framework showed relatively uniform downscaling performance across the NDVI-derived quartiles, with average PSNR values ranging narrowly from 36.56 to 37.94 dB. Quartile 4 showed the highest mean PSNR, whereas the remaining quartiles displayed highly similar averages and distributional ranges. Despite these small visual differences, the Kruskal–Wallis test indicated significant variation in performance among quartiles (H(3) = 51.99, p < 0.001). Variability across quartiles was consistently low, with coefficients of variation of approximately 6.9% (Q1), 5.86% (Q2), 5.58% (Q3), and 6.70% (Q4). Overall, greener vegetation conditions appeared to enhance reconstruction accuracy when utilizing the Bicubic interpolation approach. Under the cross-sensor framework, Bicubic performance also showed relatively uniform mean PSNR values across quartiles, with averages of 17.52 dB (Q1), 20.08 dB (Q2), 20.01 dB (Q3), and 19.95 dB (Q4). Although the quartiles exhibited similar means, their PSNR ranges (Figure 8a) revealed a clear ascending trend, indicating improvements under higher NDVI conditions. As in the intra-sensor case, the Kruskal–Wallis test confirmed significant differences among quartiles (H(3) = 117.10, p < 0.001). Variability was higher than in the intra-sensor framework but remained relatively consistent across quartiles, with coefficients of variation of approximately 16.11% (Q1), 14.76% (Q2), 15.79% (Q3), and 17.29% (Q4). Overall, Bicubic downscaling performance was relatively stable across NDVI classes, with slightly improved reconstruction under greener, higher-NDVI conditions. This suggests that even simple statistical interpolation methods benefit from stronger spectral signals associated with denser vegetation.

3.1.2. Comparison of Downscaling Strategy Frameworks (NDVI-Based Quartiles)

Model performance varied substantially across the three downscaling strategy frameworks for all evaluated models (SRGAN, ESRGAN, and Bicubic; Figure 12). Kruskal–Wallis tests confirmed significant differences among frameworks: SRGAN: H(2) = 1879.7, p < 0.001; ESRGAN: H(2) = 2652.8, p < 0.001; Bicubic: H(1) = 1543.6, p < 0.001. Across models, PSNR values consistently declined from the intra-sensor framework to the cross-sensor and intra-to-cross generalization settings, reflecting the increasing difficulty of transferring representations across heterogeneous sensor domains. While bicubic interpolation achieved slightly higher PSNR in the intra-sensor case due to its deterministic reconstruction, its performance degraded sharply under cross-sensor conditions. In contrast, GAN-based approaches (particularly ESRGAN) maintained more stable performance across frameworks, suggesting a greater capacity to learn transferable spatial representations beyond simple interpolation.

3.1.3. Comparison of Downscaling Algorithms

Across evaluation frameworks, clear differences emerged among the tested downscaling algorithms (Figure 11). Bicubic interpolation produced competitive results in the intra-sensor scenario but exhibited limited robustness when applied across sensors. Among the learning-based approaches, ESRGAN consistently outperformed SRGAN, showing higher PSNR values and reduced variability, particularly under cross-sensor and intra-to-cross generalization conditions. These results indicate that the enhanced adversarial training and feature reconstruction mechanisms of ESRGAN provide greater stability when applied to heterogeneous vegetation conditions.

3.1.4. Overall Downscaling Performance Across NDVI-Based Quartiles

The analysis highlights the influence of vegetation vigor, as captured by NDVI, on the performance of the different downscaling approaches. SRGAN exhibited a clear decline in accuracy with increasing NDVI, reflecting reduced stability in highly vegetated areas. In contrast, ESRGAN maintained greater robustness, with reconstruction accuracy improving modestly from low- to high-NDVI quartiles, suggesting that it can better capitalize on the stronger vegetation signal present in dense canopies. The Bicubic model also displayed a gradual improvement across NDVI quartiles, although its gains were more limited compared to ESRGAN. Collectively, these results highlight that model performance is not uniform across vegetation health conditions.

Figure 13 and Figure 14 shows a representative sample for visual comparison of downscaling performance for vegetation health Quartile 1 (lowest NDVI condition) and Quartile 4 (high NDVI condition), respectively, across the three strategy frameworks: intra-sensor, cross-sensor, and intra-to-cross generalization. Columns show the low-resolution (LR) input, bicubic interpolation, SRGAN and ESRGAN super-resolution predictions, and the high-resolution (HR) target reference. For each row, predictions were evaluated against the same HR target at identical spatial location. The bicubic model is not shown for the intra-to-cross framework because it does not support cross-domain generalization. Representative samples for visual comparison of downscaling performance for all quartiles are presented in the Supplementary Material (Figures S6–S9). Visual comparisons of all images in the test set (257) can be found in the Zenodo repository (https://doi.org/10.5281/zenodo.19697212) [47].

3.2. Downscaling Performance Across Structure-Derived Landscape Classes

The analysis builds on the three structure-derived landscape clusters (C1–C3) defined in the methodology, which represent a gradient of spatial heterogeneity. For comparability across models and strategies, all evaluations were conducted using a balanced set of testing tiles (n = 102). Figure 15 illustrates comparative boxplots of PSNR values across structure-derived landscape clusters for the SRGAN, ESRGAN, and Bicubic models, grouped by downscaling strategy frameworks (intra-sensor, cross-sensor, and intra-to-cross).

Table 5. Mean, standard deviation, and coefficient of variation in PSNR and SSIM values across Structured-based clusters (C1–C3), reported for each algorithm (SRGAN, ESRGAN, and Bicubic) and by downscaling strategy framework (intra-sensor, cross-sensor, and intra-to-cross generalization).

	SR GAN Model									ESR GAN Model									BICUBIC Model
Metric	Intra-Sensor			Cross-Sensor			Generalization			Intra-Sensor			Cross-Sensor			Generalization			Intra-Sensor			Cross-Sensor
	C1	C2	C3	C1	C2	C3	C1	C2	C3	C1	C2	C3	C1	C2	C3	C1	C2	C3	C1	C2	C3	C1	C2	C3
PSNR Mean	32.63	33.39	30.04	27.55	26.86	26.57	16.81	22.26	14.89	32.80	32.66	32.39	27.43	27.70	26.79	18.18	21.02	20.08	33.97	34.72	35.19	17.69	22.31	21.54
PSNR SD	1.53	2.35	7.65	1.48	3.12	2.10	2.98	2.84	8.38	1.67	2.25	1.82	1.35	2.88	2.26	2.87	3.15	4.05	2.06	2.62	2.19	1.56	3.85	3.97
PSNR CV	4.7	7.0	25.5	5.4	11.6	7.9	17.7	12.7	56.3	5.1	6.9	5.6	4.9	10.4	8.4	15.8	15.0	20.1	6.1	7.6	6.2	8.8	17.3	18.4
SSIM Mean	0.75	0.80	0.79	0.68	0.73	0.72	0.60	0.69	0.57	0.74	0.75	0.76	0.67	0.70	0.68	0.54	0.60	0.56	0.80	0.82	0.85	0.55	0.65	0.60
SSIM SD	0.06	0.08	0.07	0.05	0.08	0.06	0.08	0.08	0.18	0.06	0.08	0.06	0.05	0.08	0.06	0.08	0.09	0.13	0.06	0.07	0.05	0.06	0.10	0.10
SSIM CV	7.8	9.5	9.3	7.1	10.9	7.8	13.4	11.7	32.1	8.3	11.1	8.2	7.8	11.5	9.3	14.7	14.9	22.7	8.0	8.0	6.0	11.6	14.9	16.4

C1: Cluster 1, C2: Cluster 2, C3: Cluster 3.

shows the mean, standard deviation, and coefficient of variation (CV) values of PSNR and SSIM across Structured-based clusters. Figure S4 of the Supplementary Material illustrates comparative boxplots of SSIM values.

3.2.1. Cross-Clusters Downscaling Performance

Downscaling performance across the structure-derived clusters showed clear differences depending on the model–framework combination; therefore, results are presented separately for each case.

For SRGAN under the intra-sensor framework, model performance showed significant differences in PSNR across clusters (H(2) = 8.29, p = 0.016). Mean PSNR values were 32.63, 33.39, and 30.04 dB for C1, C2, and C3, respectively, reflecting a clear performance decline from sparse to dense woody structure. Variability also increased markedly (CV = 4.7%, 7.5%, and 25.5%), with C3 showing a wide spread of PSNR values and numerous extreme outliers. Although the mean PSNR for C3 remained near 30 dB, the presence of highly degraded predictions (<10 dB) indicates reduced model stability for dense, aggregated woody patterns. Under the cross-sensor framework, mean PSNR values were more uniform (27.55, 26.86, and 26.57 dB for C1–C3), yet cluster differences remained significant (H(2) = 18.75, p < 0.001) and dispersion increased toward C3. These patterns suggest that cross-sensor domain shifts amplify the model’s sensitivity to structural complexity. The intra-to-cross generalization framework produced the most unstable results (H(2) = 85.61, p < 0.001), with high variability across clusters (CV = 17.7%, 12.7%, and 56.3%) and numerous predictions below 10 dB, particularly in C3. This highlights the combined difficulty introduced by both dense woody aggregation and the transfer from intra-sensor training to cross-sensor testing. Collectively, the results demonstrate that SRGAN performance deteriorated as woody vegetation shifted toward dense and contiguous patterns, an effect amplified under domain-transfer conditions. This structural sensitivity may highlight the limitations of SRGAN when reconstructing high-frequency spatial detail in highly aggregated woody environments.

For the ESRGAN model under the intra-sensor framework, downscaling performance was consistent across the three structural clusters, with mean PSNR values tightly grouped between 32.39 and 32.80 dB and coefficients of variation of 5.1–6.9%. The Kruskal–Wallis test indicated no significant differences among clusters (H(2) = 5.98, p = 0.05), suggesting that structural heterogeneity exerted minimal influence when training and testing remained sensor-consistent. For ESRGAN under the cross-sensor framework, model performance maintained similar mean PSNR across clusters (26.79–27.70 dB), though variability increased progressively from C1 to C3. Statistical testing revealed significant cluster effects (H(2) = 18.31, p < 0.001), with more pronounced outliers in C3, indicating that cross-sensor transfer heightens the model’s sensitivity to structural complexity. The intra-to-cross generalization scenario yielded the weakest overall performance, with mean PSNR ranging from 18.18 to 21.02 dB and substantially higher variability (CV = 15.0–20.1%). Cluster differences were highly significant (H(2) = 40.30, p < 0.001). Although C1 exhibited lower average PSNR, C3 showed the greatest dispersion, underscoring its elevated reconstruction difficulty. Overall, ESRGAN performs poorest in Cluster 3, characterized by large, contiguous woody patches with high canopy closure, indicative of mature or late-stage woody encroachment. This structurally complex cluster consistently challenged the model—most notably under cross-sensor and intra-to-cross generalization conditions, where domain shifts further degraded reconstruction quality. Even in the intra-sensor case, the borderline significance (p = 0.05) suggests that performance degradation may emerge as structural complexity intensifies.

For Bicubic interpolation, intra-sensor performance was relatively high and consistent across clusters, with mean PSNR values of 33.97, 34.72, and 35.19 dB for C1–C3 and moderate variability (CV ≈ 6–8%). Although cluster differences were statistically significant (H(2) = 15.27, p < 0.001), the trend showed a slight increase in PSNR with increasing woody aggregation. In contrast, cross-sensor performance declined substantially, with lower mean PSNR values (17.69, 22.31, and 21.54 dB for C1–C3) and markedly higher variability (CV = 8.8–18.4%). These differences were highly significant (H(2) = 95.25, p < 0.001) and indicate that bicubic interpolation is particularly sensitive to domain shifts and structural heterogeneity. Overall, Bicubic interpolation exhibits small performance gains with increasing woody structure but reduced stability, particularly under cross-sensor transfer where structural complexity amplifies variability and weakens reliability.

3.2.2. Framework Comparison

When results were aggregated across structural clusters, clear differences emerged among the three downscaling strategy frameworks (Figure 16). Kruskal–Wallis tests confirmed significant differences for all models (SRGAN: H(2) = 626.3, ESRGAN: H(2) = 747.1, Bicubic: H(2) = 459.5; all p < 0.001). Across algorithms, PSNR consistently declined from intra-sensor to cross-sensor and further to intra-to-cross generalization frameworks, highlighting the challenges introduced by sensor heterogeneity and domain transfer. While learning-based models partially mitigated this degradation, bicubic interpolation showed the strongest performance collapse under cross-sensor transfer. These results indicate that domain shifts between sensors represent a primary limitation for remote-sensing image downscaling in structurally heterogeneous grassland environments.

3.2.3. Downscaling Algorithm Comparison

When comparing downscaling algorithms across frameworks and structural clusters (Figure 15), distinct differences in robustness emerged. Bicubic interpolation produced the highest PSNR values in the intra-sensor scenario but showed the strongest degradation under cross-sensor transfer. In contrast, GAN-based models demonstrated greater adaptability to heterogeneous sensor conditions. Among them, ESRGAN consistently outperformed SRGAN, yielding higher average PSNR and reduced variability across clusters and frameworks. SRGAN exhibited unstable behavior in several scenarios, including extreme PSNR failures (<5 dB) under intra-to-cross generalization. These results indicate that ESRGAN provides the most reliable reconstruction across structurally complex grassland landscapes.

Overall, results indicate that woody vegetation structure exerts a systematic influence on downscaling performance. Reconstruction quality generally declined as landscapes transitioned from sparse and fragmented woody patterns (Cluster 1) toward dense, aggregated configurations (Cluster 3). This effect was amplified under cross-sensor and intra-to-cross generalization frameworks, where domain shifts further degraded model performance. Among the evaluated approaches, ESRGAN showed the greatest resilience to these combined challenges, whereas SRGAN and bicubic interpolation were more strongly affected by dense woody aggregation. These findings highlight the importance of accounting for landscape structural heterogeneity when developing super-resolution approaches for grassland remote sensing. Representative samples for visual comparison of downscaling performance for all clusters are presented in the Supplementary Material (Figures S10–S12). Visual comparisons of all images in the test set (102 images) can be found in the Zenodo repository [47].

3.3. Downscaling Performance Across Texture Gradients Based on Entropy-Derived Quartiles

The analysis builds on the four entropy-based texture quartiles (Q1–Q4) defined in the methodology, capturing a gradient from low textural disorder and sparse vegetation patterns (Q1) to highly complex and spatially heterogeneous landscapes (Q4). To ensure comparability across models and strategy frameworks, evaluations were performed using a balanced testing set (n = 257). Figure 17 displays the distribution of PSNR values for the three downscaling algorithms (SRGAN, ESRGAN, and Bicubic) across the three strategy frameworks and four texture-entropy quartiles.

Table 6. Mean, standard deviation, and coefficient of variation in PSNR and SSIM values across Entropy-based quartiles, reported for each algorithm (SRGAN, ESRGAN, and Bicubic) and by downscaling strategy framework (intra-sensor, cross-sensor, and intra-to-cross generalization).

Metric	SR GAN Model												ESR GAN Model												BICUBIC Model
	Intra-Sensor				Cross-Sensor				Generalization				Intra-Sensor				Cross-Sensor				Generalization				Intra-Sensor				Cross-Sensor
	Q1	Q2	Q3	Q4	Q1	Q2	Q3	Q4	Q1	Q2	Q3	Q4	Q1	Q2	Q3	Q4	Q1	Q2	Q3	Q4	Q1	Q2	Q3	Q4	Q1	Q2	Q3	Q4	Q1	Q2	Q3	Q4
PSNR Mean	30.99	32.76	31.53	22.48	24.41	25.71	24.03	23.40	21.63	19.10	21.98	13.56	30.44	31.78	30.24	32.62	24.84	26.22	24.98	26.12	20.65	18.31	20.67	21.45	32.07	33.53	32.87	34.14	21.95	18.77	21.46	21.33
PSNR SD	2.07	2.09	1.96	11.85	2.46	1.91	2.34	3.52	3.00	1.93	3.48	8.53	2.18	1.92	1.81	2.51	2.45	1.65	1.91	2.24	2.80	2.29	3.41	3.51	2.41	2.25	2.19	2.60	2.59	2.47	3.38	3.57
PSNR CV	6.7	6.4	6.2	52.7	10.1	7.4	9.8	15.0	13.9	10.1	15.8	62.9	7.2	6.0	6.0	7.7	9.8	6.3	7.7	8.6	13.6	12.5	16.5	16.4	7.5	6.7	6.7	7.6	11.8	13.2	15.7	16.7
SSIM Mean	0.69	0.77	0.73	0.71	0.62	0.68	0.63	0.60	0.66	0.64	0.67	0.56	0.67	0.74	0.70	0.76	0.60	0.66	0.61	0.65	0.61	0.56	0.61	0.65	0.75	0.79	0.79	0.82	0.58	0.56	0.56	0.60
SSIM SD	0.08	0.07	0.07	0.15	0.07	0.06	0.07	0.11	0.08	0.07	0.08	0.17	0.09	0.07	0.07	0.08	0.08	0.06	0.06	0.07	0.08	0.07	0.08	0.09	0.08	0.07	0.06	0.06	0.10	0.07	0.09	0.10
SSIM CV	12.1	8.6	9.1	20.9	11.9	8.6	10.7	19.2	12.3	10.2	12.4	30.6	13.1	9.5	9.9	10.2	12.9	9.2	10.3	11.2	12.5	13.0	13.5	14.4	10.3	8.7	7.4	7.4	16.6	13.4	16.3	16.5

shows the mean, standard deviation, and coefficient of variation (CV) values of PSNR and SSIM across Entropy-based quartiles. Figure S5 of the Supplementary Material illustrates comparative boxplots of SSIM values.

3.3.1. Downscaling Model Performance Across Entropy-Based Texture Quartiles

The performance of the three downscaling models (SRGAN, ESRGAN, and Bicubic) varied across entropy-based texture quartiles and across the three strategy frameworks (intra-sensor, cross-sensor, and intra-to-cross generalization). Because each model–framework combination exhibited distinct behaviors in relation to textural complexity, results are presented separately for each case.

For the SRGAN model under the intra-sensor strategy framework, PSNR values exhibited pronounced quartile-dependent variation (H = 124.11, p < 0.001). Mean PSNR remained relatively consistent for Q1–Q3 (30.99–32.76 dB) but dropped sharply in Q4 to 22.48 dB. The set of boxes in Figure 17 for this model-strategy setting clearly reinforces this pattern: Q4 shows the tallest whiskers, numerous extreme PSNR values even lower than 10 dB), and the largest coefficient of variation (≈53%), reflecting reduced stability in highly disordered tonal environments. This finding may suggest that SRGAN cannot preserve predictive stability under very high textural complexity, even when both training and testing data come from the same sensor. For the SRGAN model under the cross-sensor strategy framework, quartile differences were less pronounced in magnitude, yet remained statistically significant (H = 107.51, p < 0.001). Mean PSNR values were more compact (23.44–25.71 dB), but the boxplot shows a moderately wider dispersion toward Q4, consistent with coefficients of variation increasing from 7–8% (Q1–Q2) to >14% in Q4. Although cross-sensor SRGAN avoids the severe collapse observed in intra-sensor Q4, it still exhibits reduced robustness, suggesting that transferring information across sensors attenuates—but does not eliminate—instability under highly complex textures. For the SRGAN model under the intra-to-cross generalization framework, quartile differences were pronounced: Q1–Q3 showed comparable PSNR mean values (21.63, 19.10, 21.98 dB), but Q4 dropped severely to 13.56 dB. Variability increased consistently from Q1 (CV = 13.9%) to Q4 (CV = 62.9%). The Kruskal–Wallis test strongly supported these differences (H = 244.84, p < 0.001). These results demonstrate that when SRGAN is trained only with UAV-derived low/high-resolution pairs, its generalization to Planet textures is highly sensitive to entropy. Thus, SRGAN exhibits severe limitations in its ability to generalize to Planet imagery with high textural disorder, reflecting restricted transferability across both sensors and complexity gradients.

For the ESRGAN model under the intra-sensor framework, performance across quartiles remained relatively uniform, with mean PSNR values of 30.44–32.62 dB and low dispersion (CV = 6–8%). Although statistical differences were detected (H = 192.59, p < 0.001), no sharp decline toward Q4 was observed. ESRGAN thus demonstrated greater resilience to entropy-induced heterogeneity in intra-sensor settings. Conversely, there was a slight trend towards an improvement in the model’s performance as it progressed from Q1 to Q4. For the ESRGAN model under the cross-sensor framework, mean PSNR values ranged narrowly (24.84–26.22 dB) with moderate variability (CV = 6.3–9.8%). Significant differences across quartiles were confirmed (H = 103.23, p < 0.001), though visually the distributions were relatively stable, with no abrupt Q4 degradation. Entropy effects were present but less severe than in SRGAN. For the ESRGAN model under the intra-to-cross generalization framework, quartiles showed considerable dispersion and moderately high variability (CV = 12–17%), with mean PSNR values of 20.65, 18.31, 20.67, and 21.45 dB. The Kruskal–Wallis test indicated significant quartile differences (H = 150.0, p < 0.001). Although variability remained elevated, a slight upward shift in means toward Q4 suggests mild robustness against high-entropy textures.

For the Bicubic model under the intra-sensor framework, performance showed a gradual improvement on PSNR values from Q1 (32.07 dB) to Q4 (34.14 dB), with moderate dispersion across quartiles (CV = 6.7–7.6%). Significant quartile differences were detected (H = 103.68, p < 0.001). Despite its simplicity, Bicubic displayed stable behavior with no high-entropy degradation. For the Bicubic model under the cross-sensor framework, mean PSNR values fluctuated across quartiles (21.95 → 18.77 → 21.46 → 21.33 dB), while variability increased progressively (CV = 11.8–16.7%). Quartile differences were statistically significant (H = 174.13, p < 0.001). The increasing dispersion toward Q4 suggests that entropy exacerbates sensitivity to spectral inconsistencies in cross-sensor conditions.

3.3.2. Comparison of Downscaling Strategy Frameworks (Entropy-Based Quartiles)

When results were aggregated across entropy quartiles, clear differences emerged among the three downscaling strategy frameworks (Figure 18). For all models, PSNR decreased systematically from the intra-sensor to the cross-sensor and further to the intra-to-cross generalization framework (SRGAN: H(2) = 1477.0; ESRGAN: H(2) = 2380.6; Bicubic: H(1) = 1530.2; p < 0.001). This trend reflects the increasing difficulty of transferring representations across heterogeneous sensor domains. Bicubic interpolation achieved competitive performance in the intra-sensor setting but showed pronounced degradation under cross-sensor transfer. In contrast, GAN-based models maintained higher PSNR values in cross-sensor scenarios, suggesting that adversarial learning enables the extraction of spatial representations that generalize more effectively across sensor domains.

3.3.3. Comparison Among Downscaling Algorithms

When comparing downscaling algorithms across frameworks and entropy quartiles (Figure 17), clear differences in robustness emerged. Bicubic interpolation achieved the highest PSNR values under the intra-sensor framework (33.15 dB), indicating strong performance when training and testing data originate from the same sensor. However, its performance degraded substantially under cross-sensor conditions. In contrast, GAN-based models demonstrated greater adaptability to sensor heterogeneity, with ESRGAN achieving the highest PSNR under cross-sensor conditions (25.54 dB). Under intra-to-cross generalization, ESRGAN also outperformed SRGAN in both mean PSNR and stability, whereas SRGAN exhibited extreme variability, including occasional failures with PSNR below 5 dB. Overall, ESRGAN showed the greatest robustness across entropy gradients and sensor-transfer scenarios, while SRGAN was particularly sensitive to highly disordered textures. These results highlight the importance of selecting downscaling algorithms capable of maintaining stable performance across heterogeneous sensors and complex texture environments.

Across the three landscape descriptors examined (NDVI-based vegetation condition, structure-derived woody aggregation clusters, and entropy-based texture quartiles) downscaling performance exhibited consistent sensitivity to increasing landscape heterogeneity. In general, reconstruction quality declined as scenes transitioned from spectrally uniform and structurally sparse conditions toward environments characterized by dense woody aggregation and high textural disorder. These effects were particularly evident under cross-sensor and intra-to-cross generalization frameworks, where domain shifts amplified the influence of landscape complexity on model stability. Among the evaluated algorithms, ESRGAN demonstrated the greatest robustness across all heterogeneity gradients, maintaining relatively stable PSNR values even under high entropy and dense woody configurations. In contrast, SRGAN exhibited pronounced instability in highly complex scenes, while bicubic interpolation performed competitively only under intra-sensor conditions but degraded markedly under sensor transfer. Collectively, these findings highlight that landscape heterogeneity—whether expressed through vegetation condition, spatial structure, or tonal complexity—represents a fundamental constraint for remote-sensing image downscaling and should be explicitly considered when developing and evaluating super-resolution models.

Representative samples for visual comparison of downscaling performance for all Entropy-derived quartiles are presented in the Supplementary Material (Figures S13–S16). Visual comparisons of all images in the test set (257 images) can be found in the Zenodo repository [47].

4. Discussion

This study demonstrates that landscape heterogeneity and operational context jointly shape the performance of image downscaling models in grassland ecosystems, extending and contextualizing insights from prior remote sensing and ecological research. Consistent with landscape ecology theory, our results show that spectral condition, spatial configuration, and textural disorder influence the predictive value of coarse-resolution imagery and, consequently, the capacity of super-resolution algorithms to reconstruct fine-scale patterns [3,38]. This landscape dependence contrasts with the implicit assumption in many super-resolution studies that model performance is uniform across conditions [23,29,39], highlighting the limitations of global, dataset-averaged metrics for evaluating ecological remote sensing tasks.

Three consistent patterns emerge: (i) model performance is not uniform across landscape conditions, (ii) robustness to heterogeneity differs markedly among downscaling strategies and architectures, and (iii) operational domain shifts often amplify landscape-driven degradation, particularly in complex environments.

4.1. Influence of Vegetation Condition on Downscaling Performance

Vegetation condition, as characterized by NDVI gradients, exerted a strong and model-dependent influence on super-resolution fidelity. SRGAN showed pronounced sensitivity to increasing vegetation density, with reconstruction quality degrading and variability increasing under greener conditions. This behavior suggests that high-NDVI grasslands—characterized by dense canopies and elevated spectral variability—challenge SRGAN’s ability to recover stable high-frequency detail. This decline in performance aligns with the known limitations of the standard MSE-based and initial adversarial losses in SRGAN, which often struggle to recover stable high-frequency details in stochastic textures such as dense canopies [30].

In addition, SRGAN incorporates Batch Normalization, which can introduce artifacts and instability when input statistics differ from the training distribution, particularly in high-NDVI regions [48]. The standard perceptual loss formulation may also be insufficient to capture fine-scale stochastic textures, leading to oversmoothing or inconsistent detail reconstruction [49].

In contrast, ESRGAN demonstrated substantially greater robustness across NDVI gradients, with stable or modestly improving performance under higher vegetation vigor. This finding supports prior evidence that enhanced perceptual and adversarial losses in ESRGAN reduce texture loss and improve structural consistency [28]. ESRGAN removes Batch Normalization and introduces improved feature extraction and perceptual loss formulations, enhancing its ability to reconstruct complex high-frequency textures and reducing the artifacts observed in SRGAN [49]. The observed improvement under greener conditions suggests that dense vegetation provides richer and more coherent spatial patterns that ESRGAN can exploit more effectively than SRGAN as reported by Toosi [50]. Bicubic interpolation, while competitive under sensor-consistent conditions, showed limited adaptability and offered only marginal gains with increasing NDVI, reinforcing its reliance on local smoothness rather than learned spatial structure.

Collectively, these results demonstrate that vegetation vigor is not merely a background condition but an active determinant of downscaling success, challenging the assumption that GAN-based models perform uniformly across ecological gradients consistent with prior studies [32].

4.2. Role of Landscape Structure and Spatial Configuration

Landscape structure emerged as a dominant control on downscaling performance, particularly under conditions of dense and aggregated woody cover. Across all models, reconstruction quality declined as spatial patterns transitioned from sparse and fragmented to highly contiguous and clumped configurations. This effect was most pronounced for SRGAN, which exhibited instability and failure modes in structurally complex landscapes, especially when combined with cross-sensor transfer. ESRGAN again showed greater resilience, maintaining comparatively stable performance across structural clusters, although degradation became evident under strong domain shifts. These findings indicate that while advanced GAN architectures can partially mitigate the challenges posed by structural heterogeneity, dense woody aggregation represents a fundamental limit for super-resolution reconstruction in grasslands. Such patterns are dominated by abrupt transitions, shadowing effects, and high-frequency spatial contrasts that are difficult to infer reliably from coarse-resolution inputs. This finding is consistent with ecological observations that aggregated patches and high-canopy closure reduce the distinct characteristics of spectral patterns within coarse pixels [1,15], thereby constraining any algorithm’s ability to disaggregate fine-scale information.

Bicubic interpolation showed limited sensitivity to structural complexity under intra-sensor settings but suffered marked instability under cross-sensor application, underscoring its lack of transferability. Together, these results highlight that spatial configuration—not only spectral content—plays a critical role in governing downscaling fidelity and that structurally complex grasslands pose persistent challenges across methods.

4.3. Effects of Textural Complexity and Entropy Gradients

Textural complexity, quantified through entropy gradients, exerted one of the strongest controls on model robustness. High-entropy landscapes consistently degraded downscaling performance, particularly for SRGAN, which showed severe instability and loss of fidelity under highly disordered textures. This behavior suggests that SRGAN struggles to distinguish meaningful spatial patterns from noise when textural cues are weak or highly variable, leading to unreliable reconstructions. This outcome strengthens earlier suggestions that textural complexity influences the resolution enhancement tasks [23,29,38].

ESRGAN exhibited markedly greater tolerance to textural disorder, maintaining stable performance and avoiding catastrophic degradation even under high-entropy conditions. This resilience likely reflects its enhanced capacity to preserve perceptual structure while suppressing spurious texture generation. Bicubic interpolation, although stable in sensor-consistent settings, became increasingly variable under cross-sensor conditions, indicating that entropy amplifies spectral inconsistencies that simple interpolation cannot resolve.

These findings confirm that textural complexity is also a critical—but often overlooked—dimension in evaluating super-resolution performance and that entropy-based stratification provides valuable insight into model limitations that are obscured by global performance metrics.

4.4. Impact of Downscaling Strategy Frameworks and Domain Shifts

Consistent with concerns raised in the recent literature on domain shifts in remote sensing applications [40,41], our results demonstrate that downscaling strategy frameworks strongly influence reconstruction fidelity. Across all stratification schemes, downscaling strategy frameworks exerted a systematic and dominant influence on performance. Reconstruction accuracy consistently declined from intra-sensor to cross-sensor settings and further under intra-to-cross generalization, reflecting the increasing difficulty of transferring learned representations across sensor domains. This pattern highlights the central role of distribution shift in constraining real-world applicability of super-resolution models. This pattern corroborates previous reports that naive application of models trained in one sensor domain may yield misleading results in another due to spectral and contextual differences [1,15].

This domain mismatch is likely caused by differences in the spectral response functions of the MicaSense RedEdge-P and PlanetScope sensors. Variations in band placement, bandwidth, and radiometric sensitivity influence how surface reflectance is measured, which may result in discrepancies in spectral signatures even for the same target [51,52]. These discrepancies propagate through the learning process, restricting the transferability of features learnt under one sensor setup to another.

Although convolutional neural networks typically learn representations tied to the statistics of their training data, evaluating cross-sensor scenarios remains important for quantifying the practical limits of super-resolution models in multi-sensor remote-sensing applications. In many operational contexts, high-resolution UAV imagery is spatially limited, whereas satellite sensors provide the broader coverage required for landscape-scale monitoring.

While bicubic interpolation performed competitively under intra-sensor conditions, its performance deteriorated sharply under cross-sensor transfer, emphasizing its limited operational utility beyond controlled settings. GAN-based approaches, particularly ESRGAN, demonstrated superior adaptability and retained meaningful reconstruction capability under domain shifts, although performance losses remained unavoidable. These findings suggest that domain-adaptation approaches, such as transfer learning or limited fine-tuning with target-sensor data, may help bridge UAV-based training datasets with broader satellite observations.

The combined effect of landscape heterogeneity and domain shift showed a consistent amplification of performance degradation across conditions, suggesting a non-additive interaction between these factors, although this relationship is inferred from observed patterns rather than formally quantified through interaction or variance decomposition analysis.

4.5. Implications for Ecological Applications and Model Deployment

Taken together, these results highlight that the reliability of super-resolution products in grassland ecosystems depends not only on model architecture but also on landscape context and deployment strategy. ESRGAN emerges as the most robust option across heterogeneous conditions, particularly when cross-sensor generalization is required. However, even advanced GAN-based models exhibit limitations in landscapes characterized by dense woody aggregation and extreme textural disorder.

From an applied perspective, these findings suggest that landscape-aware evaluation should be a prerequisite for operational deployment of super-resolution models in ecological monitoring. In homogeneous or low-structure environments, simpler methods may suffice, whereas heterogeneous and structurally complex landscapes require carefully selected architectures and realistic expectations regarding reconstruction fidelity. More broadly, this study demonstrates that incorporating ecological heterogeneity into model assessment is essential for bridging the gap between methodological innovation and reliable real-world application of GAN-based super-resolution.

This study is not intended as a comprehensive model benchmark, but as a controlled analysis of how landscape heterogeneity and domain shift influence super-resolution behavior. While focused on grassland ecosystems, the proposed framework is transferable to other heterogeneous landscapes and can be extended to additional model families in future work.

4.6. Limitations and Future Research Directions

The present study was designed as a controlled landscape-aware evaluation rather than an exhaustive benchmark of all super-resolution architectures. Accordingly, SRGAN and ESRGAN were selected as representative GAN-based models to examine how ecological heterogeneity and sensor-domain shifts influence performance under comparable conditions. While this design enabled clear attribution of performance differences to landscape and transfer effects, it does not imply that GANs fully represent the broader super-resolution literature. Future studies should extend this framework to include non-adversarial approaches such as residual CNN models, transformer-based architectures, and diffusion models in order to determine whether the heterogeneity patterns identified here are model-specific or more universal. Likewise, the present analysis focused on grassland ecosystems, whose diffuse vegetation–soil mixtures and fine-scale transitions provide a distinct but not exclusive form of spatial heterogeneity. Validation across forests, deserts, croplands, and other landscapes would help assess the broader transferability of these findings. Finally, this study prioritized reconstruction-fidelity metrics (PSNR and SSIM) to quantify pixel-level and structural agreement under controlled transfer scenarios; however, perceptual and distribution-based metrics such as FID may provide complementary insight into visual realism and should be incorporated in future evaluations. Collectively, these directions would broaden the applicability of the proposed landscape-aware framework while preserving its central objective: identifying when environmental complexity and domain mismatch constrain super-resolution reliability.

5. Conclusions

This study presents the first landscape-aware assessment of GAN-based super-resolution models in natural grassland ecosystems, explicitly integrating vegetation condition, spatial structure, and textural complexity with operational downscaling frameworks. Our results reveal that landscape heterogeneity substantially mediates model performance. SRGAN showed marked sensitivity to dense vegetation, highly aggregated woody structure, and high textural entropy, with performance declining and variability increasing under these conditions. In contrast, ESRGAN maintained robust and stable performance across NDVI, structure, and texture gradients, suggesting that its adversarial and perceptual learning mechanisms enable more effective extraction of transferable spatial features. Bicubic interpolation performed competitively in intra-sensor scenarios but lacked adaptability to cross-sensor and domain-shift conditions, reinforcing the limitations of conventional interpolation approaches in heterogeneous grassland landscapes.

These findings highlight the critical role of ecological context in guiding the application of super-resolution methods. By linking model performance to NDVI, patch structure, and texture, this study provides practical insight into the conditions under which different downscaling approaches remain reliable in heterogeneous grasslands. More broadly, the results emphasize that performance evaluation of deep learning-based remote sensing methods must explicitly account for ecological heterogeneity and operational domain shifts. Rather than serving as a benchmark comparison of architectures, this study proposes a framework for understanding how landscape complexity governs super-resolution behavior in natural ecosystems.