Terrain-Dependent Effects of SAR Speckle Filtering on Land Cover Classification Using Sentinel-1

Abstract

Synthetic aperture radar (SAR) data from Sentinel-1 enable land cover classification independent of cloud cover and illumination; however, classification performance is affected by inherent speckle noise. This study evaluates the influence of eight speckle filtering algorithms on classification accuracy using Sentinel-1 Ground Range Detected (GRD) data across five contrasting terrain types in eastern Slovakia (mountain, forest, urban, cropland, and water). Speckle suppression was assessed using Peak Signal-to-Noise Ratio (PSNR), Mean Squared Error (MSE), Structural Similarity Index (SSIM), and Equivalent Number of Looks (ENL). Classification performance was quantified using Support Vector Machine (SVM), Random Forest (RF), and Histogram-based Gradient Boosting (HistGB) under VV, VH, and dual-polarization (VV + VH) configurations with repeated balanced sampling. Classification accuracy varies across terrain types. In croplands, Lee Sigma combined with SVM in VV + VH mode achieved Overall Accuracy (OA) = 0.746 ± 0.010, whereas in mountainous areas, OA = 0.838 ± 0.005 was achieved with Intensity-Driven Adaptive Neighborhood (IDAN) filtering. Urban areas achieved OA = 0.890 ± 0.006, whereas forest classification remained limited (best OA = 0.582 ± 0.011). Water surfaces approached saturation accuracy (OA ≈ 0.9998). Dual polarization improved performance in heterogeneous environments but had a limited effect in homogeneous classes. The results show that terrain structure influences the interaction between speckle filtering and classification performance.

1. Introduction

Synthetic aperture radar (SAR) has become an established component of Earth observation [1] for land surface monitoring [2] due to its capability to operate independently of solar illumination and atmospheric conditions [3]. Unlike optical sensors, which are limited by cloud cover and daylight availability, SAR systems provide consistent data acquisition under all-weather conditions and during both day and night [4]. This property makes SAR particularly valuable for applications requiring temporal stability and operational reliability [5], including flood monitoring [6], agricultural assessment [7], forest mapping [8], urban expansion analysis [9], and water resource management [10].

The Sentinel-1 mission within the European Copernicus programme has significantly expanded the accessibility of SAR data. The twin satellites Sentinel-1A and Sentinel-1C provide freely available dual-polarization imagery (VV and VH) at approximately 10 m spatial resolution with regular revisit intervals. These characteristics enable large-scale and regional analyses [11] and support the integration of SAR data into land cover classification workflows [12]. In recent years, Sentinel-1 imagery has been increasingly used in combination with machine learning techniques [13] for thematic mapping and environmental monitoring [14].

Despite its operational advantages, SAR imagery is inherently affected by speckle noise [15]. Speckle arises from the coherent nature of radar signal acquisition and manifests granular intensity fluctuations in the image [16]. Unlike additive noise in optical imagery, speckle is multiplicative and depends on the local backscatter amplitude [17]. This phenomenon reduces radiometric stability [18], lowers local contrast, and may obscure fine spatial structures [19]. Consequently, speckles directly affect the performance of classification algorithms [20], especially in tasks that rely on pixel-level discrimination and texture-based features [21].

To mitigate speckle effects, numerous filtering algorithms have been developed over the past decades [22]. Classical non-adaptive approaches, such as Boxcar or Median filtering, apply simple statistical operations within a fixed moving window [23]. While computationally efficient, these methods often introduce spatial blurring and degrade edge information [24]. In contrast, adaptive filters including Lee, Lee Sigma, Frost, Gamma Maximum A Posteriori (Gamma MAP), and Intensity-Driven Adaptive Neighborhood (IDAN) incorporate local statistics or structural constraints to balance variance reduction and detail preservation [25]. These methods are designed to reduce speckle noise while maintaining radiometric fidelity and spatial gradients [26].

Most comparative evaluations of speckle filters focus on radiometric quality indicators. Metrics such as Peak Signal-to-Noise Ratio (PSNR), Mean Squared Error (MSE), Structural Similarity Index (SSIM), and Equivalent Number of Looks (ENL) are commonly used to quantify variance suppression and structural preservation [27]. Such analyses provide insight into statistical behavior in homogeneous and heterogeneous regions [28]; however, improved radiometric metrics do not necessarily imply improved thematic classification performance [29]. A filter that maximizes ENL, for example, may simultaneously suppress discriminative spatial features required for class separability [30].

In parallel, the application of Sentinel-1 SAR data to land cover classification [31] has expanded considerably, often employing supervised machine learning methods [32] such as Support Vector Machines (SVMs) [33], Random Forests (RFs) [34], and gradient boosting (XGBs) [35] algorithms. In many of these studies, speckle filtering is treated as a standard preprocessing step, and a single filter (frequently Lee or Refined Lee) [22] is adopted without a systematic comparison against alternative strategies. Furthermore, classification experiments are typically conducted within a single landscape context or under homogeneous environmental conditions [36], which limits the transferability of conclusions to structurally contrasting terrains.

The combined effects of filtering strategy [37], terrain structure [38], polarization mode (VV and VH) [39], and acquisition geometry [40] have not been systematically evaluated within a unified benchmarking design. Variations in terrain are associated with different scattering regimes [41]: surface scattering dominates over smooth water and uniform agricultural fields [42], double-bounce interactions are common in built-up areas [43], and volume scattering characterizes forested canopies [44]. These scattering differences influence backscatter variability and consequently the way filtering alters local contrast and class separability [45]. Dual-polarization input may improve discrimination by incorporating additional scattering information [46], although its contribution is not uniform across land cover types [47]. Differences between ascending and descending passes modify incidence angle [48] and viewing geometry [49], which can affect backscatter distribution and classification stability [50], particularly in areas with pronounced topographic relief.

Existing literature rarely integrates these interacting factors within a unified experimental design. Few studies simultaneously (i) evaluate multiple filtering algorithms [51], (ii) compare single and dual-polarization configurations [52], (iii) account for acquisition geometry, and (iv) assess classification performance using repeated balanced sampling to control for stochastic variability in machine learning models [40]. As a result, it remains unclear whether filters that optimize radiometric homogeneity [53] also maximize thematic accuracy across structurally heterogeneous environments.

Addressing this gap is relevant for operational SAR-based land cover mapping [54], where preprocessing decisions can directly influence classification of robustness and thematic reliability [55]. Understanding the terrain-dependent behavior of speckle filtering is particularly important for single-date Sentinel-1 applications, where temporal averaging is not available to reduce speckle variance.

This study aims to systematically evaluate the influence of eight speckle filtering algorithms on land cover classification accuracy using Sentinel-1 Ground Range Detected (GRD) data across five contrasting terrain types in eastern Slovakia: mountainous terrain (High Tatras), forested uplands (Čergov Mountains), urban environment (Košice), agricultural lowland (near Trebišov), and a water body (Zemplínska Šírava reservoir). These areas represent a wide spectrum of structural complexity, elevation range, and dominant scattering mechanisms.

In addition to comparing filtering strategies, the study investigates the influence of polarization configuration (VV, VH, and combined VV + VH), orbit geometry (ascending and descending acquisitions), and window size in the Lee Sigma filter. Radiometric filtering performance is evaluated using PSNR, MSE, SSIM, and ENL metrics. Classification performance is assessed using Support Vector Machine, Random Forest, and Histogram-based Gradient Boosting classifiers within a repeated balanced sampling framework to ensure statistical comparability across configurations.

The study addresses the following research questions:

• Q1: How does the choice of speckle filtering algorithm affect land cover classification accuracy across different terrain types?
• Q2: Does the optimal filtering strategy vary with land cover heterogeneity and dominant scattering mechanisms?
• Q3: How do polarization configuration and orbit direction influence classification robustness in the presence of speckle filtering?

This study combines radiometric assessment with terrain-based classification experiments to examine how speckle suppression influences thematic accuracy. The results contribute to a better understanding of preprocessing effects in SAR-based land cover mapping. Different effects of speckle filtering can be expected across terrain types. Larger improvements are likely in more heterogeneous areas such as cropland and mountains, while only minor changes are expected in homogeneous surfaces such as water. Forest areas remain more challenging due to higher variability, and urban areas are expected to show moderate improvements.

2. Study Area

For this study, five areas of interest were selected in eastern Slovakia, representing diverse types of land cover and geomorphological conditions. The selection criteria included variations in vegetation type and density, elevation range, hydrological features, and the presence of urban structures. The study focuses on areas selected for their relevance in testing speckle noise reduction algorithms and performing subsequent land cover classification using Sentinel-1 radar imagery. To capture different imaging geometries, the analysis includes two Sentinel-1 scenes: one from an ascending orbit and the other from a descending orbit. Geospatial coverage of images is indicated in the map (Figure 1) by blue rectangles. Red rectangles denote the specific AOIs extracted for detailed analysis.

Figure 1. Overview map of the study areas used for SAR speckle noise filtering evaluation, showing the spatial distribution of five test sites in eastern Slovakia representing different land cover types: (a) mountainous area, (b) forested area, (c) urban area, (d) cropland area, and (e) water body area. Blue boxes indicate the Sentinel-1 scenes used in the analysis. Detailed zooms illustrate the extent and landscape context of each site.

The following list provides a detailed description of the selected AOIs, marked on the map (Figure 1), which were extracted from the radar images and further analyzed. Selection criteria focused on specific SAR characteristics, including backscatter intensity, surface homogeneity, texture, and sensitivity to geometric effects. These properties have a significant impact on the performance of speckle filters and the outcomes of subsequent classification.

• High Tatras—High Mountain Environment
  The Tatra region represents extreme morphological conditions characterized by significant elevation differences, rocky landforms, alpine vegetation, and sparse forest cover. In this environment, the SAR signal is strongly affected by incidence geometry and shadowing, creating suitable conditions for testing the robustness of different filtering algorithms and their impact on image classification.
• Čergov Mountains—Forested Area
  This area features continuous deciduous and coniferous forest cover with moderately rugged terrain typical of the flysch Carpathian Mountains. From the radar imaging perspective, it represents a complex environment with heterogeneous backscatter structures, allowing performance evaluation of speckle filters in areas with gentle slopes and moderate elevation.
• Košice—Urban Area
  The urban area of Košice exemplifies a highly reflective and strong SAR backscatter environment caused by multipath propagation and structural reflections from buildings and infrastructure. This location is suitable for testing classification accuracy in environments with high textural complexity.
• East Slovak Lowland near Trebišov—Cropland
  This area consists of intensively used agricultural land with predominantly flat terrain and regular surface structure. Due to its homogeneous texture and low SAR signal dispersion, it provides an ideal setting for testing SAR data processing in low-noise environments.
• Zemplínska Šírava Reservoir—Water Surface
  This area represents a homogeneous water surface with very low SAR backscatter and minimal local elevation variability at the water surface, although the broader AOI includes surrounding terrain with elevations ranging from 100 to 443 m. Its relative radiometric uniformity makes it an appropriate environment for evaluating speckle filtering performance, particularly using the ENL metric.

Table 1 summarizes the key parameters and site characteristics of all AOIs, including their geographical and topographical attributes of all investigated areas.

Table 1. Basic characteristics of the selected study areas representing different land cover types in eastern Slovakia, including geographic location, elevation, and topographic complexity.

AOI	Land Cover Type	Location Name	Area (km2)	Centroid Coordinates (WGS 84)	Elevation (m a.s.l.)	Max Slope (°)
(a)	Mountains	High Tatras Mountains	157.358	49.20758° N 20.11755° E	964–2655	82.3
(b)	Forests	Čergov Mountains	66.653	49.24208° N 21.01043° E	521–1149	47.2
(c)	Urban	Košice Urban Area	74.476	48.71357° N 21.25542° E	186–496	36.7
(d)	Croplands	East Slovak Lowland near Trebišov	160.542	48.80655° N 21.72004° E	102–388	23.8
(e)	Water	Zemplínska Šírava Reservoir	97.663	48.78707° N 22.01349° E	100–443	36.5

The purpose of selecting these sites was to cover the full spectrum of contrasting environments from a SAR imaging perspective, ranging from homogeneous and smooth surfaces (such as water bodies and croplands) to complex and structurally challenging areas (including urban regions and mountainous terrains). This selection enables a comprehensive assessment of the robustness and versatility of various speckle filtering algorithms and qualitative metrics, depending on the type of area and the SAR backscatter characteristics.

3. Materials and Methods

3.1. Materials

This study uses Sentinel-1 Ground Range Detected (GRD) radar images acquired in Interferometric Wide Swath (IW) mode. The data were provided by the European Space Agency (ESA) under the Copernicus Programme and accessed through the Copernicus Open Access Hub. The dataset comprises two distinct acquisition geometries: an ascending pass on 7 October 2024 and a descending pass on 11 October 2024, which provide different radar incidence angles that help reduce geometric distortions and minimize shadowing or layover effects in the areas of interest (Table 2).

Table 2. Acquisition parameters of Sentinel-1A GRD imagery used in this study, including orbital information, imaging mode, and polarization.

Satellite Type	Acquisition Date	Orbit Direction	Absolute Orbit	Beam Mode	SAR Polarization
Sentinel-1A	7 October 2024	Ascending	55,999	IW	VV, VH
Sentinel-1A	11 October 2024	Descending	56,050	IW	VV, VH

GRD products contain radar backscatter intensity in two polarizations: VV (vertically transmitted, vertically received) and VH (vertically transmitted, horizontally received). The spatial resolution of the images is approximately 10 m, with a pixel spacing of 10 × 10 m. A single IW mode scene covers a swath width of approximately 250 km.

The selection of acquisition dates considered both meteorological and phenological conditions in the study areas. October corresponds to a period of reduced vegetation activity, which decreases the influence of volume scattering from dense canopies and enables a more accurate representation of surface structural characteristics in the radar signal. This period is also characterized by relatively stable atmospheric conditions with no extreme precipitation or snow cover, thereby minimizing variability in soil moisture and water surface conditions between acquisitions. The combination of ascending and descending passes provides a more comprehensive representation of radar backscatter from different incidence angles. It is essential for evaluating speckle filter performance and subsequent classification.

At the same time, the two acquisitions differ slightly in local incidence angles due to ascending and descending viewing geometry, which may influence the backscatter response, particularly in areas with complex topography. Although the scenes were acquired within a short time interval (four days), minor variations in surface conditions, such as soil moisture or crop state, cannot be fully excluded, especially in cropland and forest areas.

The use of VV and VH polarizations aimed to enhance the capability to differentiate between various land cover types. VV polarization is particularly sensitive to reflections from smooth surfaces and to specular reflection at oblique incidence angles (e.g., water bodies, bare soils, infrastructure) [56], allowing for the detection of differences in surface roughness and orientation relative to the radar beam. VH polarization, representing a cross-polarization mode, is more sensitive to volume scattering, which is characteristic of vegetated areas or complex structures where multiple signal reflections occur [57]. The combination of both polarizations enables more robust separation of vegetation from non-vegetated surfaces [58], supports the detection and delineation of water bodies, and reduces the risk of misclassification under heterogeneous surface conditions.

3.2. Methods

Sentinel-1 time series data support monitoring of temporal changes in land surface properties. In this study, the focus is on two acquisition instances (ascending and descending) to evaluate the effects of speckle filtering parameters, polarization, and acquisition geometry on classification performance. The aim is methodological rather than temporal. The methodology chapter presents subsections on SAR image preprocessing, speckle filtering algorithms, image classification, quantitative assessment of filtering performance, and accuracy metrics. For a clearer overview of the workflow, Figure 2 shows main steps.

Figure 2. Workflow for SAR-based land cover classification, including data acquisition, preprocessing, speckle filtering, feature extraction, and machine learning classification and evaluation. Colors are used to distinguish the individual processing stages.

3.2.1. Preprocessing SAR Images

For the analysis, the images underwent pre-processing according to the standard ESA chain, which consisted of several key steps [59]. The workflow began with orbital correction to ensure accurate pixel geolocation. Thermal noise removal was then applied, and radiometric calibration transformed radar signal amplitudes into physically meaningful backscatter coefficient values (σ⁰). The application of suitable filtering methods suppressed speckle noise while preserving both edges and homogeneous areas. Orthorectification, based on the SRTM 1 Arc-second digital elevation model, corrected geometric distortions caused by terrain. Finally, the data was reprojected into the WGS 84/UTM Zone 34N coordinate system, allowing for seamless integration with other spatial layers and enabling precise analysis. The pre-processed images thus provide a solid basis for applying various speckle noise reduction methods, described in the following subsection.

3.2.2. SAR Speckle Filtering Algorithms

Speckle filters are applied to reduce SAR noise while preserving structural details and edges [60]. It allows accurate representation of features and improves the visual quality of the image. The method involves moving a kernel across the image. At each position, the technique performs calculations and updates the central pixel to smooth the local area. Several filters, developed for this purpose, are suitable for different noise characteristics.

The filters fall into two main categories. The first comprises non-adaptive filters, which apply simple statistical operations within a fixed window (e.g., Boxcar or Median filter) [61]. The second includes adaptive filters, which consider local image statistics or structure and adjust computations according to the properties of the region (e.g., Lee, Lee Sigma, Refined Lee, Frost, Gamma MAP, and IDAN) [62]. Each algorithm employs a different approach to modeling noise and signal. Non-adaptive filters are computationally efficient but tend to blur edges more strongly. Adaptive filters adjust their calculations based on local image statistics or structure, offering a more effective trade-off between noise reduction and the preservation of fine details. The following section summarizes the filters employed in this study, along with their fundamental principles and mathematical formulas.

• Boxcar filter represents the simplest approach, based on the arithmetic mean of pixel intensities within a fixed moving window [63]. This filter effectively reduces noise in homogeneous regions; however, it causes edge blurring and loss of structural details.

  $$\widehat{I}={\displaystyle \frac{1}{N}}{\displaystyle \sum _{i=1}^N}{I}_i$$

  (1)

  where $I_i$ denotes the intensity values within the window and $N$ is the number of pixels in the kernel.
• Median filter is a nonlinear filter based on the median value within a predefined neighborhood [64]. It replaces the central pixel with the median of all intensity values inside the kernel:

  $$\widehat{I}=median(I_1, I_2,\dots ,I_N)$$

  (2)

  Its main advantage lies in suppressing extreme values (i.e., “salt-and-pepper” noise) and better edge preservation compared to the Boxcar filter. However, it is generally less effective for multiplicative SAR speckle noise compared to adaptive filters.
• Lee filter is an adaptive filter specifically designed for SAR imagery, assuming a multiplicative noise model [65]. The filtered pixel value is estimated as a weighted combination of the local mean and the original pixel value:

  $$\widehat{I}=\overline{I}+W(I-\overline{I})$$

  (3)

  where $\bar{I}$ represents the local mean and $W$ is a weighting factor derived from the ratio of local variance to noise variance.
• Lee Sigma filter extends the Lee filter by incorporating a sigma-based thresholding mechanism [66]. Only pixels whose intensities fall within a predefined range around the local mean (typically within ±$k\sigma$) are included in the filtering process:

  $$Iϵ\left[\overline{I}-k\sigma ,\overline{I}+k\sigma \right]$$

  (4)

  where $\overline{I}$ is the local mean, $\sigma$ is the local standard deviation, and $k$ is the sigma threshold parameter. This selective averaging strategy enhances edge preservation and prevents excessive smoothing in heterogeneous regions.
• Refined Lee filter is an advanced extension of the Lee filter that incorporates directional windows and gradient-based operations. The dominant local gradient direction is first determined, and filtering is then performed adaptively along structural features, thereby reducing object blurring [67].

  $$\widehat{I}={\overline{I}}_{dir}+W_{dir}(I-{\overline{I}}_{dir})$$

  (5)

  This approach significantly improves the preservation of edges and linear features, making the Refined Lee filter particularly suitable for complex landscapes such as urban environments and mountainous terrain.
• Frost filter is an adaptive exponential filter that assigns weights to neighboring pixels based on both spatial distance and local variance [68]:

  $$\widehat{I}={\displaystyle \sum _{i=1}^N}{I}_i·e^{-\alpha d_i}$$

  (6)

  where $d_i$ is the distance from the central pixel and $\alpha$ is a damping factor related to local variance. This filter enables effective speckle suppression while partially preserving edges, although some loss of fine texture may still occur.
• Gamma MAP (Gamma Maximum A Posteriori) filter is a grounded approach based on Bayesian estimation, assuming a Gamma distribution for SAR backscatter [69]. The filtered pixel value is obtained by maximizing the posterior probability:

  $$\widehat{I}=\arg {max}_I\left[p\left(I{I}_{obs}\right)\right]$$

  (7)

  The Gamma MAP filter performs well in homogeneous regions and preserves radiometric consistency (e.g., for water body detection). However, its effectiveness may decrease in highly heterogeneous scenes.
• IDAN (Intensity-Driven Adaptive Neighborhood) filter dynamically defines the filtering window based on local radiometric similarity rather than a fixed spatial kernel [70]. Pixels are included in the adaptive neighborhood only if their intensities are statistically similar to that of the central pixel:

  $$\widehat{I}={\displaystyle \frac{1}{M}}\sum _{iϵ{\Omega }_{IDAN}}{I}_i$$

  (8)

  where ${\Omega }_{IDAN}$ denotes the adaptive neighborhood and $M$ is the number of selected pixels. This approach is effective in homogeneous areas; however, its adaptive nature may lead to inconsistent smoothing in highly textured regions.

3.2.3. Quantitative Assessment of SAR Speckle Filtering

Image quality degradation in SAR imagery is primarily caused by speckle noise, blurring, and other unwanted artifacts introduced during acquisition and processing. The objective evaluation of filtering performance relies on quantitative metrics that assess both noise suppression and the preservation of structural image information. When a distortion-free reference image is available, full-reference metrics are applied; otherwise, statistical measures derived directly from the processed image are used. In this study, the performance of speckle reduction algorithms was evaluated using Mean Squared Error (MSE), Peak Signal-to-Noise Ratio (PSNR), Equivalent Number of Looks (ENL), and Structural Similarity Index Measure (SSIM) [71]. The original (unfiltered) SAR image is used as a reference for full-reference metrics. Although it still contains speckle, this approach enables a relative comparison between filtering strategies under identical conditions.

• Mean Squared Error (MSE) quantifies the average squared difference between the original (reference) image and the filtered image. It reflects the overall radiometric deviation introduced by processing and is defined as:

  $$MSE={\displaystyle \frac{1}{MN}}{\displaystyle \sum _{i=1}^M}{\displaystyle \sum _{j=1}^N}{\left(X_{í,j}-X_{i,j}^{\prime }\right)}^2$$

  (9)

  where $X_{í,j}$ is the pixel value in the original image, $X_{i,j}^{\prime }$ is the pixel value in the filtered image, and $M\times N$ is the image size. Lower MSE values indicate smaller deviations and therefore higher similarity between images [71].
• Peak Signal-to-Noise Ratio (PSNR) expresses the ratio between the maximum possible signal intensity and the distortion represented by MSE. It is measured in decibels (dB) and calculated as:

  $$PSNR=10 {log}_{10}\left({\displaystyle \frac{{\left(2^n-1\right)}^2}{MSE}}\right)=10 {log}_{10}\left({\displaystyle \frac{{255}^2}{MSE}}\right)$$

  (10)

  where $n$ is the number of bits per pixel. Higher PSNR corresponds to better image quality, reflecting more effective noise suppression relative to signal strength [72].
• Equivalent Number of Looks (ENL) was computed as a no-reference indicator of speckle suppression in homogeneous regions. ENL represents the effective number of independent intensity samples averaged per pixel and is defined as:

  $$ENL={\displaystyle \frac{mean {\left(F\right)}^2}{var \left(F\right)}}$$

  (11)

  where $F(x,y)$ denotes the filtered image. Higher ENL values indicate a more effective improvement of the signal-to-noise ratio in homogeneous regions, corresponding to better image quality [72].
• Structural Similarity Index Measure (SSIM) is employed to assess the similarity between two images. This metric assesses the similarity between two images by analyzing their luminance, contrast, and structural information. It provides a more accurate estimate of visual distortions as perceived by the human eye compared to simpler metrics, such as MSE. The index is calculated as follows:

  $$SSIM={\displaystyle \frac{\left(2{\mu }_x{\mu }_y+C_1\right) \left(2{\sigma }_{xy}+C_2\right)}{\left({\mu }_x^2+{\mu }_y^2+C_1\right) \left({\sigma }_x^2+{\sigma }_y^2+C_2\right)}}$$

  (12)

  where ${\mu }_x$ and ${\mu }_y$ are the mean values of images $x$ and $y$, ${\sigma }_x^2$ and ${\sigma }_y^2$ are the variances, ${\sigma }_{xy}$ is the covariance, $C_1$ and $C_2$ are constants introduced to prevent instability when denominators are small:

  $$C_1={\left(k_{1}L\right)}^2$$

  (13)

  $$C_2={\left(k_{2}L\right)}^2$$

  (14)

  here $L$ represents the dynamic range of pixel values, and the default values are $k_1=0.01$, $k_2=0.03$. The resulting SSIM value ranges from −1 to 1, where 1 corresponds to identical images and thus the highest quality [73].

It should be noted that the original SAR image is not a noise-free reference, as it already contains speckle noise. For this reason, PSNR and MSE are used here only for relative comparison between filtering methods, not as absolute measures of image quality.

3.2.4. SAR Image Classification

To quantify the impact of speckle filtering on land-cover discrimination in Sentinel-1 SAR imagery, supervised classification was performed in a controlled benchmarking setup. For each acquisition date and speckle-filtered product, identical reference masks, sampling strategy, feature definitions, and classifier settings were used to ensure comparability across filters. The experiment was formulated as an AOI-specific binary classification task (one-versus-background) for five land-cover types: Urban, Water, Cropland, Forest, and Mountain. For each area of interest (AOI), pixels were labeled as Target (belonging to the given class) or NonTarget (background). This formulation minimizes contextual class mixing and enables a direct evaluation of separability changes induced by different speckle filtering approaches (Table 3).

Table 3. Experimental design and data configuration used for classification.

Component	Specification
Classes	Urban, Water, Cropland, Forest, Mountain
Reference Data	ESA WorldCover 2021 (v200), Copernicus DEM GLO-30
Sampling	Balanced (2000 train + 2000 test per class)
Polarizations	VV + VH, VV, VH
Texture Window	5 × 5
Runs	5
Classifiers	SVM (RBF), RF (600 trees), HistGB (300 iterations)

Reference labels for Urban, Water, Cropland, and Forest were derived from ESA WorldCover 2021 (v200). The following class codes were used: Forest (10), Cropland (40), Built-up/Urban (50), and Water (80). The dataset was accessed via STAC and resampled to the Sentinel-1 grid using nearest-neighbor resampling to preserve categorical integrity. The Mountain class was defined using a terrain-based proxy derived from Copernicus DEM GLO-30. The DEM was reprojected to the SAR grid (bilinear resampling), and pixels were labeled as Mountain where elevation ≥ 900 m and slope ≥ 12°, with slope computed directly on the target grid. To reduce boundary uncertainty and mixed-pixel effects, binary masks were morphologically eroded prior to sampling. A default erosion depth of two pixels was applied for most AOIs, while cropland masks were eroded by three pixels to minimize contamination from field boundaries. Additionally, for the Cropland AOI, the NonTarget class was restricted to structurally distinct WorldCover categories (Forest, Built-up, Water, Wetlands) to reduce ambiguity and evaluate cropland separability against clearly contrasting backgrounds.

Training and testing datasets were generated by random sampling without replacement from the eroded masks using a balanced design. For each AOI, 2000 Target and 2000 NonTarget pixels were used for training (4000 samples total), and an additional 2000 Target and 2000 NonTarget pixels were used for independent testing. The train/test ratio was therefore 1:1, and class imbalance was avoided by construction. Sampling was repeated over five independent runs (N_RUNS = 5) using different random seeds. Within each run, identical spatial sample locations were reused across all speckle-filtered datasets, ensuring that observed differences in classification performance were attributable to filtering effects rather than sampling variability. The samples were distributed across the AOIs to reduce the influence of local spatial autocorrelation, although some degree of spatial dependence cannot be completely avoided.

Classification features were derived from speckle-filtered Sentinel-1 sigma0 backscatter in three polarization configurations: VV + VH (combined VV and VH), VV-only, and VH-only. For each sampled pixel, intensity values were extracted along with texture descriptors computed in a 5 × 5 moving window, namely the local mean and standard deviation for each polarization. All feature extraction was performed on co-registered rasters. Higher-order texture features (e.g., GLCM) were not included to isolate the effect of speckle filtering on basic intensity-derived descriptors.

Three non-linear classifiers commonly applied in remote sensing were evaluated: Support Vector Machine (SVM) with radial basis function kernel, Random Forest (RF), and Histogram-based Gradient Boosting (HistGB). The SVM was implemented as a pipeline including z-score standardization (StandardScaler) followed by an RBF kernel classifier with C = 10.0, gamma = “scale”, and class_weight = “balanced”. The Random Forest classifier used 600 trees, max_features = “sqrt”, class_weight = “balanced”, and a run-specific random seed with parallel processing enabled. The HistGradientBoosting classifier was configured with learning_rate = 0.1, max_depth = 8, and max_iter = 300 (number of boosting iterations), with the random seed set per run. These algorithms were selected as robust and widely adopted baselines for SAR-based land-cover classification. Tree-based ensemble methods (RF and HistGB) are resilient to noise and capable of modeling non-linear feature interactions, while SVM with RBF kernel is well suited for high-dimensional and non-linearly separable SAR feature spaces.

For each AOI, polarization mode, speckle filter, and run, classifiers were trained on the training subset and evaluated on the independent test subset. The resulting predictions were used to compute standard accuracy statistics, which are described in the subsequent subsection.

3.2.5. Accuray Assessment Metrics

Classification performance was evaluated on an independent test set using complementary accuracy metrics that quantify overall correctness, agreement beyond chance, and class-wise discrimination. For each AOI, polarization mode, speckle filter, classifier, and run, predicted labels were compared to reference labels, and a confusion matrix was compiled. In the binary (Target vs. NonTarget) setting, the confusion matrix entries correspond to true negatives (TN), false positives (FP), false negatives (FN), and true positives (TP), enabling the derivation of standard performance indicators.

• Overall Accuracy (OA) was computed as the proportion of correctly classified samples among all test samples:

  $$OA={\displaystyle \frac{TP+TN}{TP+TN+FP+FN}}$$

  (15)

  OA provides an intuitive summary of classification correctness but can be insensitive to class-specific errors if class proportions are imbalanced.
• Cohen’s Kappa coefficient ($\kappa$) was additionally calculated to address potential chance agreement effects. Kappa measures agreement between predictions and reference labels corrected for agreement expected by random chance:

  $$\kappa ={\displaystyle \frac{p_o-p_e}{1-p_e}}$$

  (16)

  where $p_o$ is the observed agreement (equivalent to OA) and $p_e$ is the expected agreement computed from the marginal class frequencies of the confusion matrix. Values of $\kappa$ range from −1 to 1, with 1 indicating perfect agreement and 0 representing agreement equivalent to chance.
• F1-score (OA) represents the harmonic mean of precision and recall:

  $$F1={\displaystyle \frac{2·Precision·Recall}{Precision+Recall}}$$

  (17)

  where

  $$Precision={\displaystyle \frac{TP}{TP+FP}} Recall={\displaystyle \frac{TP}{TP+FN}}$$

  (18)
• Macro-averaged F1-score (Macro-F1) was reported as an unweighted mean of F1-scores computed separately for the Target and NonTarget classes:

  $${F1}_{Macro}={\displaystyle \frac{{F1}_{Target}+{F1}_{NonTarget}}{2}}$$

  (19)

  Macro-F1 emphasizes balanced performance across classes and is therefore particularly informative when the objective is to preserve discrimination of both the Target class and its background rather than optimizing accuracy for the majority class.

To account for sampling variability, the classification experiment was repeated across multiple runs with different random seeds. For each configuration (AOI × polarization mode × filter × classifier), metrics were summarized as the mean and standard deviation across runs, providing a robust estimate of expected performance and its stability.

4. Results

This section presents the outcomes of speckle filtering and its influence on SAR image quality and subsequent land cover classification accuracy. Both visual and quantitative assessments were conducted across different filtering methods, window sizes, polarizations, and orbit directions. Additionally, post-classification land cover area statistics are evaluated to understand the practical implications of filtering on thematic mapping.

4.1. Visual Assessment of Speckle Filtering

Differences in speckled noise attenuation and spatial detail were observed among the applied filters. The original Sentinel-1 scenes showed strong granular backscatter fluctuations typical of SAR data, most evident over homogeneous targets such as water bodies and cropland. This noise severely reduced visual interpretability and obscured fine spatial features, especially in low-backscatter regions (Figure 3).

Figure 3. Illustrative comparison of speckle filtering techniques for Sentinel-1 VH polarization over an “Urban” AOI.

Boxcar and Median filtering reduced local backscatter variability; however, this reduction was accompanied by spatial averaging. Edge transitions became less distinct, and fine-scale structures in urban areas were partially suppressed. While these filters improved image homogeneity, excessive spatial averaging at larger window sizes led to over smoothing in heterogeneous landscapes.

Adaptive speckle filters, including Lee, Refined Lee, Lee Sigma, Frost, Gamma Map, and IDAN, showed a favorable balance between noise suppression and the preservation of spatial features [74]. Speckle variance decreased across all tested filters, while the scene’s spatial variability was largely retained. Lee Sigma and Refined Lee preserved edge gradients [64] in urban structures and maintained intra-class variability within forest stands. Frost and Gamma Map produced stronger attenuation in areas with low backscatter variability, most evident over water surfaces, where the signal became spatially uniform without a substantial shift in mean intensity.

Differences between the polarization channels were observable in the filtered data. The VV channel displayed relatively consistent backscatter patterns across the analyzed AOIs. The VH channel showed increased variability in vegetated areas, reflecting the contribution of volume scattering and associated speckle. These visual trends suggest that the effectiveness of speckle suppression is strongly dependent on both land cover characteristics and polarization configuration [75].

The visual comparison indicates that adaptive filters maintained spatial structures to a greater extent while reducing speckle-related fluctuations. In contrast, non-adaptive smoothing approaches introduced more pronounced spatial averaging, which affected structural representation in heterogeneous areas.

4.2. Quantitative Image Quality Metrics After SAR Filtering

Quantitative evaluation of the filtered SAR images was performed using PSNR, MSE, SSIM, and ENL metrics, enabling a comprehensive assessment of speckle suppression and the preservation of structural image properties across different land-cover types (Table 4). The numerical results are derived from the VV polarization, while Figure 4 provides an overview for the same metrics using VH polarization acquired on the same date.

Table 4. Table of the comparison of eight applied SAR speckle filtering algorithms across five monitored locations, based on the Sentinel-1 image acquired on 7 October 2024 in VV polarization.

AOI	Metrics	SAR Speckle Filter Algorithm
		Boxcar	Median	Lee	Lee Sigma	Refined Lee	Frost	Gamma MAP	IDAN
(a) Mountains	PSNR	22.24	23.70	22.24	18.43	22.66	23.29	22.24	18.86
	MSE	388.47	277.18	288.33	933.72	352.78	304.65	388.36	844.87
	SSIM	0.85	0.88	0.85	0.60	0.88	0.87	0.85	0.61
	ENL	5.77	5.66	5.76	6.65	6.34	5.75	5.77	6.66
(b) Forests	PSNR	22.86	22.62	22.88	17.64	21.76	23.09	22.88	16.98
	MSE	331.07	347.72	330.15	1077.27	416.34	311.7035	329.65	1302.31
	SSIM	0.87	0.88	0.87	0.56	0.88	0.88	0.87	0.54
	ENL	2.73	2.68	2.73	3.69	2.91	2.75	2.74	2.93
(c) Urban	PSNR	19.58	21.56	19.59	14.56	19.44	21.25	19.58	16.75
	MSE	709.88	450.03	709.77	2239.52	727.67	484.02	709.88	1373.38
	SSIM	0.84	0.87	0.84	0.52	0.85	0.87	0.84	0.57
	ENL	8.66	7.57	8.66	14.91	10.25	8.25	8.66	11.71
(d) Croplands	PSNR	24.12	24.25	24.13	18.92	23.30	24.83	24.13	18.69
	MSE	249.88	242.38	249.47	813.51	299.58	213.62	249.32	879.91
	SSIM	0.85	0.87	0.82	0.52	0.88	0.87	0.85	0.45
	ENL	5.05	4.91	5.05	7.33	5.40	5.05	5.05	6.25
(e) Water	PSNR	24.98	25.34	24.99	19.81	24.36	25.83	24.99	19.60
	MSE	206.50	190.17	206.28	679.76	238.23	169.99	206.20	712.71
	SSIM	0.89	0.91	0.89	0.65	0.91	0.91	0.89	0.43
	ENL	5.01	6.79	4.99	8.63	6.62	6.66	5.01	7.26

Figure 4. Heatmaps of speckle filtering performance across five land-cover types based on Sentinel-1 VH-polarized imagery acquired on 7 October 2024. Metrics include PSNR, MSE, SSIM, and ENL. Rows represent filtering algorithms and columns correspond to areas of interest (AOIs). Blue outlines indicate the best-performing filter for each AOI and metric. For MSE, the best-performing filter corresponds to the lowest value. Similar values across filters should be interpreted as comparable rather than strictly different.

Across all analyzed locations, applying speckle filters resulted in a marked increase in ENL values compared to unfiltered images, indicating effective noise reduction. The highest ENL values were observed primarily for the adaptive Lee Sigma and IDAN filters, particularly in urban areas (ENL up to 14.91) and agricultural landscapes (ENL up to 7.33), where more homogeneous surfaces allow greater stabilization of the backscatter signal. In forested areas, ENL values were generally lower (approximately 2.7–3.7), reflecting structural complexity and dominant volume scattering [76]. High ENL values reflect strong variance reduction, but this does not always translate into better preservation of spatial detail or improved classification performance.

The Median and Frost filters achieved the highest PSNRs and the lowest MSEs in most AOIs. The effect was most pronounced over water bodies (PSNR up to 25.83 for Frost) and agricultural areas (24.83), where the backscatter is dominated by surface scattering and spatial variability is low. Under these conditions, local averaging reduces multiplicative speckle without substantially altering the mean signal level.

The differences between filters are also visible in Figure 4. Filters that maximize ENL (Lee Sigma, IDAN) form a distinct group with strong smoothing, while Median, Frost, and Refined Lee maintain higher structural similarity.

Lee Sigma consistently produced lower PSNR and higher MSE values. Its adaptive weighting reduces local variance but also suppresses edges and small-scale contrast, which becomes more apparent in heterogeneous areas.

A similar tendency was observed in the SSIM results. Median, Frost, and Refined Lee reached SSIM values between 0.87 and 0.91, with higher values over urban surfaces and water bodies. Lee Sigma and IDAN ranged between 0.45 and 0.65. The decrease corresponds to stronger local variance suppression and altered spatial configuration of the signal.

Metric behavior differed across land-cover types. Water and croplands reached higher PSNR, SSIM, and ENL values [22]. Forested areas showed lower values, consistent with the dominant volume scattering within vegetation canopies. Multiple scattering processes and randomly oriented scatterers increase inherent variability, limiting ENL improvement and reducing structural correspondence after filtering. This increased variability also directly affects classification performance by reducing feature separability in subsequent machine learning tasks.

The quantitative metrics show distinct behavior among the evaluated filters. Median and Frost achieved effective speckle reduction in areas dominated by surface scattering, with limited alteration of spatial patterns, which is consistent with previous findings [76]. Lee Sigma and IDAN emphasize variance reduction at the expense of local contrast, as reported in sigma-based and adaptive neighborhood filtering studies [66]. The choice of filtering approach should therefore align with the requirements of subsequent analysis, particularly in applications where class discrimination depends on subtle spatial details. In other words, filters that achieve higher ENL or PSNR do not necessarily provide better classification results, which highlights the need to consider the final application when selecting a filtering method.

4.3. Impact of Speckle Filtering on Classification Accuracy

The impact of speckle filtering on classification accuracy varied across the evaluated AOIs. The effect was most pronounced in structurally heterogeneous environments, whereas in homogeneous scenes the filtering effect was limited. To assess the absolute benefit of speckle filtering, classification performance was also evaluated using unfiltered SAR data as a baseline. The comparison focuses on representative AOIs to illustrate both absolute and relative improvements.

In cropland AOIs, the Lee Sigma–SVM combination in dual polarization (VV + VH) reached OA = 0.746 ± 0.010 with κ = 0.491 ± 0.020. IDAN in the same configuration yielded OA = 0.744 ± 0.009. Single-polarization schemes showed lower values overall. The best VH-only setup reached OA ≈ 0.691, whereas VV-only remained at approximately 0.633. The dispersion across repeated runs was small for the adaptive filters with the highest OA. Compared to the unfiltered baseline, the best-performing configuration improved OA by approximately +0.049. Compared to the Boxcar filter, the improvement was smaller (+0.008), indicating that while speckle filtering provides clear benefits over raw data, differences between filtering strategies are more limited (Figure 5).

Figure 5. Scene-dependent sensitivity of land-cover classification accuracy to speckle filtering strategy based on Sentinel-1 VV + VH. The best-performing configurations are highlighted, while small differences between filters should be interpreted with caution. Black dots indicate outlier values.

The relatively low classification performance in the forest AOI (OA ≈ 0.582) is likely related to the complex backscatter response of vegetation. In contrast to other land-cover types, forest areas produce a mixture of scattering contributions, which increases variability within the class and reduces the distinction between target and background pixels.

It is also reflected in the lower ENL values observed in forest areas (Table 4) and in slightly higher classification variability across repeated runs. These observations indicate that the variability is not solely due to speckle noise but primarily to the heterogeneous structure of the forest canopy. Although adaptive filters such as IDAN reduce speckle noise, they do not substantially alter the signal itself. As a result, their impact on classification accuracy remains limited in this context. Furthermore, the feature set used in this study relies on intensity values and simple local statistics. This representation is likely insufficient to capture the structural complexity of forest environments, potentially leading to overlapping feature distributions and lower classification performance. Improvements over the unfiltered baseline were limited (approximately +0.015), indicating that classification performance is primarily constrained by intrinsic variability rather than speckle noise.

In the mountain AOI, the highest OA was obtained with IDAN filtering combined with SVM in dual polarization (VV + VH), reaching 0.838 ± 0.005 with κ = 0.676 ± 0.011. Single-polarization inputs (VV and VH) did not achieve this level of accuracy. Across the tested filters, adaptive approaches were typically associated with higher accuracy in this terrain type. A substantial improvement was observed compared to the unfiltered baseline, with OA increasing by approximately +0.342. Compared to the Boxcar filter, the improvement was modest (+0.006), suggesting that most of the gain is achieved by speckle suppression rather than by the specific choice of filtering algorithm.

Urban areas reached OA values close to 0.90 in several configurations. Lee Sigma combined with dual polarization (VV + VH) yielded 0.890 ± 0.006. Similar accuracy levels were obtained with both SVM and Random Forest classifiers. When only a single polarization was used, OA decreased to approximately 0.863 for VV and 0.832 for VH.

Water surfaces showed near-saturated accuracy across filters and polarizations. The highest OA was recorded for Refined Lee (≈0.9998 ± 0.0002). Differences among filtering strategies and polarization modes were minimal. Heterogeneous scenes were more sensitive to the choice of filtering strategy and polarization input, with adaptive and dual-polarization setups yielding higher OA, confirmed in previous studies [77]. In homogeneous areas, the impact of filtering was limited and differences among configurations were small.

These results confirm that speckle filtering is a necessary preprocessing step for SAR-based land cover classification, particularly in heterogeneous environments, where it yields substantial improvements in classification accuracy.

4.4. Influence of Lee Sigma Window Size

The Lee Sigma filter represents an adaptive speckle filtering approach that exploits local statistical properties within a moving window to suppress speckle noise while attempting to preserve structural image content. In this experiment, the influence of three window sizes (5 × 5, 7 × 7, and 9 × 9) was evaluated using quantitative image quality metrics for two polarization configurations (VV and VH). The assessment focused on radiometric and structural indicators, including PSNR, MSE, SSIM, and ENL (Table 5).

Table 5. Comparison of Lee Sigma filter performance for different window sizes and SAR polarizations across land cover types.

AOI	Metrics	Lee Sigma Window Size
		VV Polarization			VH Polarization
		5 × 5	7 × 7	9 × 9	5 × 5	7 × 7	9 × 9
(a) Mountains	PSNR	19.32	18.43	17.42	17.93	10.64	10.57
	MSE	761.12	933.72	1176.60	1046.47	5607.06	5700.21
	SSIM	0.68	0.60	0.59	0.61	0.21	0.20
	ENL	6.36	6.65	7.98	8.13	0.70	1.71
(b) Forests	PSNR	18.50	17.64	16.24	16.96	19.84	15.98
	MSE	896.47	1077.27	1498.63	1299.74	1336.06	1641.26
	SSIM	0.58	0.56	0.35	0.56	0.54	0.34
	ENL	3.50	3.69	5.08	3.54	3.54	1.75
(c) Urban	PSNR	17.18	14.56	15.24	16.82	15.11	15.12
	MSE	1234.12	2239.52	1916.41	1353.23	1990.68	2000.47
	SSIM	0.61	0.52	0.41	0.61	0.53	0.36
	ENL	10.53	14.91	15.90	4.26	4.70	3.05
(d) Croplands	PSNR	20.16	18.92	17.89	20.39	19.99	12.07
	MSE	616.94	813.51	1031.39	594.27	651.26	4036.68
	SSIM	0.53	0.52	0.27	0.56	0.53	0.32
	ENL	7.04	7.33	10.09	2.94	3.09	3.47
(e) Water	PSNR	21.29	19.81	19.11	21.01	20.66	11.75
	MSE	482.61	679.76	799.01	514.90	558.20	4343.15
	SSIM	0.65	0.65	0.49	0.67	0.66	0.41
	ENL	5.27	8.63	13.34	3.43	2.56	5.63

For the VV polarization, the smallest window (5 × 5) achieved the highest PSNR (19.32), the lowest MSE (761.12), and the highest SSIM (0.68). Increasing the window size gradually increased ENL values (up to 7.98 for the 9 × 9 window), indicating stronger speckle suppression due to spatial averaging of backscatter intensity within a larger neighborhood. However, this improvement in noise statistics was accompanied by a decrease in SSIM, reflecting a progressive loss of local spatial variability.

Across the analyzed land-cover types, the 5 × 5 window provided the most balanced trade-off between speckle suppression and structural preservation. Although ENL consistently increased with window size, this metric alone does not capture structural fidelity. Larger windows produced stronger smoothing, which reduced structural similarity despite improved noise statistics. Under identical window configurations, the VV polarization generally achieved higher PSNR and SSIM values than the VH polarization.

The results show that window size selection significantly influences the balance between speckle suppression and structural preservation [78]. The 5 × 5 configuration providing the most favorable quantitative performance under the evaluated conditions [79].

In addition to the radiometric metrics, the influence of window size was evaluated at the classification stage. The comparison was performed for the same window configurations (5 × 5, 7 × 7, and 9 × 9) across five land-cover categories.

The results (Figure 6) show that the influence of window size on classification accuracy is moderate but differs between land-cover types, confirmed also by [80]. In cropland and mountainous areas, a slight increase in OA can be observed with larger windows. In these environments, moderate spatial averaging reduces speckle-related fluctuations in backscatter intensity, which improves the stability of class separation.

Figure 6. Impact of Lee Sigma speckle filter window size (5 × 5, 7 × 7, 9 × 9) on SAR classification accuracy for different land-cover types. Boxplots show the distribution of OA values across multiple runs, while points represent individual classification results. Differences between window sizes are generally small, and the results should be interpreted as overall trends rather than strict separations.

For forested areas, the influence of window size is minimal, and OA values remain similar across the tested configurations. This is likely related to the dominance of volume scattering within forest canopies, where the radar signal already contains substantial intrinsic variability. Urban and water classes show similarly small sensitivity to window size because these surfaces typically produce stable radar responses dominated by double-bounce (urban) or specular scattering (water).

The classification results show that moderate window sizes perform most consistently. Although larger windows increase ENL and suppress speckle more strongly, they also introduce stronger spatial smoothing, which may reduce local structural variability important for class discrimination. As shown in Figure 6, the differences in OA between the tested window sizes are relatively small and tend to stabilize for the largest configuration.

The results indicate that the relationship between speckle suppression and classification accuracy is not linear [29]. Window sizes of 5 × 5–7 × 7 provide a reasonable compromise between noise reduction and preservation of spatial information, whereas the 9 × 9 window mainly increases smoothing without a consistent improvement in classification accuracy. This behavior corresponds with the radiometric analysis, where stronger filtering improved noise statistics but reduced structural detail in the SAR signal. Overall, the analysis confirms that Lee Sigma window size affects both radiometric image quality and classification performance. Moderate window sizes (5 × 5–7 × 7) provide the most stable balance between speckle suppression and preservation of spatial information, whereas larger windows primarily increase smoothing without consistent classification benefits [81]. The relatively small variability across repeated runs suggests that the observed differences are stable, although no formal statistical testing was performed.

4.5. Comparison of VV and VH Polarization Configurations

The comparison of polarization configurations indicates that the contribution of dual-polarization features depends on the structural characteristics of the analyzed land-cover type. Across most AOIs, the VV + VH configuration achieved the highest performance [82], although the magnitude of improvement varied substantially (Figure 7).

Figure 7. Pairwise Polarization Effects on Classification Performance (Mean ΔOA and ΔMacro-F1 across Five Sampling Runs). While dual polarization generally improves performance, some differences between configurations remain small. Bold values and black frames highlight the largest absolute effects.

In croplands areas, the difference between polarization modes was the most distinct. The VV + VH configuration reached OA = 0.746 compared to 0.633 (VV) and 0.691 (VH). The mean improvement relative to VV was +0.110 in OA and +0.113 in Macro-F1. VH exceeded VV by +0.056 in OA. Agricultural surfaces typically contain vertically structured vegetation elements that produce strong volume scattering, which contribute significantly to cross-polarized returns, confirmed by other studies [83]. The combination of VV surface responses and VH volume scattering, therefore, improves class separability.

In mountainous terrain, dual polarization also yielded higher accuracy. The maximum OA (0.838) was obtained in VV + VH mode, corresponding to an increase of approximately +0.035 relative to VV. The same tendency was observed in Macro-F1 and κ. Complex terrain introduces strong variability in incidence angle, surface roughness, and vegetation cover. The combination of co-polarized and cross-polarized signals captures complementary scattering responses from bare surfaces, vegetation patches, and shadowed slopes.

Urban areas showed a similar but more asymmetric pattern. The VV + VH configuration reached OA = 0.890. VV-only and VH-only achieved approximately 0.863 and 0.832, respectively. The difference between VH and VV was −0.036. The combined VV + VH input exceeded VH by +0.070. Built-up structures typically generate strong double-bounce scattering between vertical walls and ground surfaces [84], which is primarily preserved in co-polarized channels. This suggests that cross-polarized information alone is insufficient in built-up environments but enhances separability when combined with co-polarized backscatter.

Forest classification showed limited sensitivity to polarization configuration. The best result reached OA = 0.582 ± 0.011, and pairwise differences remained below +0.005. Forest canopies are dominated by volume scattering from randomly oriented branches and leaves, resulting in substantial variability in both the VV and VH channels. Under single-date SAR conditions, the separability between forest and surrounding vegetation classes remains limited, and the additional polarization channel provides only minor improvement.

For water surfaces, classification accuracy approached saturation across all polarization modes (best OA ≈ 0.9998). Pairwise differences were negligible, indicating that separability was dominated by strong radiometric contrast rather than polarization diversity. Calm water bodies typically produce specular reflection, resulting in very low backscatter in both VV and VH channels.

Overall, dual-polarization configurations were associated with higher accuracy in cropland, mountain, and urban AOIs. In homogeneous or radiometrically dominant classes, the differences between single- and dual-polarization inputs were small. While the overall trends are consistent across AOIs, some of the observed differences between orbital configurations and polarization modes are relatively small and should be interpreted with caution, as no formal statistical testing was performed.

4.6. Satellite Orbit Geometry Influence (Ascending vs. Descending)

In addition to speckle filtering and polarization configuration, the satellite acquisition geometry (ascending versus descending orbit) can influence the SAR backscatter recorded from the surface. Differences in look direction and local incidence angle mean that the same area is not observed under identical geometric conditions. These effects are most noticeable in areas with pronounced relief, where changes in viewing geometry can modify the local radar response through geometric effects such as foreshortening, layover, or radar shadow.

The comparison between the ascending acquisition (7 October 2024) and the descending acquisition (11 October 2024) shows that orbit geometry affects classification performance differently across the analyzed AOIs. For homogeneous classes such as open water, the influence of orbit direction was negligible. Classification accuracy remained nearly identical for both acquisitions (best OA ≈ 0.9998). Calm water surfaces primarily produce specular reflection away from the radar sensor, resulting in very low backscatter in both VV and VH channels. Consequently, small differences in look direction or incidence angle have minimal impact on class separability. Under these conditions, the direction of the orbit did not noticeably affect class separability.

In croplands (Figure 8), higher accuracy was obtained for the ascending pass. Mean OA reached 0.738, while the descending acquisition dropped to 0.646 (Δ ≈ −0.092). The difference is present across runs—no change in ranking between configurations. The scene is strongly directional. Field patterns and parcel structure are not aligned the same way for both look directions. This affects the backscatter and, consequently, the classification, as also reported in studies highlighting the directional sensitivity of agricultural SAR backscatter [85].

Figure 8. Influence of Sentinel-1 Orbit Geometry (Ascending vs. Descending) on Classification Results Using SVM, RF, and HistGB (or HGB) with the IDAN Speckle Filter. Observed differences between orbit directions vary across AOIs and are in some cases minor.

Forest classification showed similarly low accuracy for both orbit directions. The best ascending and descending configurations reached comparable OA values (approximately 0.58). Forest backscatter is dominated by volume scattering from branches, trunks, and foliage, producing substantial intrinsic variability in both VV and VH channels. Under single-date SAR conditions, this volumetric scattering tends to dominate over geometric effects related to orbit direction, resulting in only minor differences between ascending and descending observations.

Urban areas show only a small shift between acquisitions. OA increased from 0.834 (ascending) to 0.846 (descending), i.e., about +0.012. The ranking of configurations stays the same. Backscatter is mainly driven by double-bounce from walls and the ground, so orientation matters. Differences between orbits are present but limited.

Mountain terrain behaved differently for the two acquisition geometries. The descending pass reached OA = 0.798, while the ascending one remained lower (0.743). The difference (~0.055) was visible for all tested polarization settings, including VV and VH. VV + VH still performed best, but the relative behavior did not change between orbits. The explanation is mainly geometric. A change in look direction alters the local incidence angle relative to slope orientation. The backscatter response is therefore uneven: slopes facing the sensor tend to appear brighter (due to foreshortening), whereas others are weakened or not recorded at all due to shadowing. In steep terrain, this behavior is spatially inconsistent, leading to irregular backscatter patterns that propagate into the classification.

Orbit geometry plays a role, but not a dominant one. The effect is uneven. It is visible in croplands and, to a lesser extent, in urban areas. In water and forest, there is almost no change. The response depends on scene structure rather than the method itself. Terrain further complicates the situation through changes in local incidence angle. These effects are not spatially consistent. Using both orbits reduces part of this variability but does not remove it completely, which aligns with recent findings on multi-geometry SAR analysis [83].

4.7. Filter and Classifier Performance Comparison

Classification results differed across AOIs and were more sensitive to filter choice than to classifier type. The contrast between filtering strategies was more evident in heterogeneous scenes, whereas homogeneous targets showed only minor variation.

In croplands, the highest accuracy was obtained with adaptive filters, particularly Lee Sigma and IDAN in VV + VH mode (OA up to ~0.75). Classical filters (Boxcar, Frost, Gamma MAP) remained below this level across all classifiers. The no-filter baseline also underperformed relative to adaptive approaches. Among classifiers, SVM reached the highest peak values, with Random Forest close behind.

Forest classification remained limited (OA ≈ 0.58) across configurations. IDAN provided the highest mean accuracy, but the differences between filters were small. Classifier choice had only a minor effect; SVM and Random Forest produced comparable results, while HistGradientBoosting was slightly lower.

In mountainous terrain, the choice of filter made a clear difference. The highest values were obtained with IDAN (OA > 0.80), while other adaptive approaches stayed close but slightly lower. Classical filters did not reach this level in any configuration. The same ordering appears across runs. SVM reached the top values, although the gap to Random Forest was small. Urban results (Figure 9) are more compact. The spread between methods is limited. Lee Sigma reached the highest OA (≈ 0.89), but several configurations fall into a similar range. Differences between filters are present, though less pronounced than in croplands or mountainous areas.

Figure 9. Comparison of best speckle filters per classification model and polarization for AOI Urban (Sentinel-1, 11 October 2024). Differences between filters are relatively small in this AOI, with several configurations achieving comparable results.

Water behaves differently. All configurations converge to almost identical results (OA ≈ 1.0). Refined Lee reaches the maximum, but the gap to other methods is negligible. Changing the filter or classifier has little practical effect here.

Across AOIs, differences between filters remain visible regardless of the classifier used. In several cases, the ordering between filters does not change, even when switching models. The classifier’s effect is present, but not consistently. Some configurations shift slightly; others do not.

5. Discussion

The results show that selecting the speckle filtering algorithm influences classification accuracy, though the effect differs across terrain types. In the heterogeneous environments analyzed here, particularly in the mountainous and cropland AOIs, adaptive filters generally produced higher accuracy than simple smoothing methods [86]. This tendency was observed across several runs, where filters such as Lee Sigma and IDAN repeatedly ranked among the best-performing configurations. Their behavior suggests that reducing speckle variance while maintaining local backscatter differences is beneficial for class discrimination. In contrast, simple averaging filters smooth spatial gradients in the radar signal, which in some cases led to slightly lower classification accuracy, even though they reduced radiometric variability. Figure 10 summarizes the relative performance of the evaluated speckle filtering algorithms across different AOIs and polarization configurations. The ranking patterns remain largely consistent between the two acquisition dates, indicating stable filter behavior across the analyzed scenes. This consistency suggests that the observed differences between filtering strategies are primarily due to terrain characteristics and scattering conditions rather than scene-specific variability, as also reported in previous studies that emphasize the scene-dependent behavior of speckle filtering performance [80].

Figure 10. Mean rank of speckle filtering algorithms across AOIs for VV, VH, and VV + VH polarization configurations. Rankings are based on overall classification accuracy (OA), where lower values indicate better performance. The mean rank was calculated from two acquisition dates (7 and 11 October 2024). Error bars show the minimum–maximum range of ranks between the two dates. Ranking differences are generally consistent, although small variations between configurations should be interpreted with caution.

Across most analyzed environments, adaptive filters ranked among the best-performing configurations. Lee Sigma and IDAN frequently produced the highest values of overall accuracy, Cohen’s κ, and Macro-F1 when dual-polarization data (VV + VH) were used as input. In cropland areas, Lee Sigma achieved an overall average accuracy of approximately 0.73 (κ ≈ 0.46), slightly outperforming IDAN. In mountainous terrain, IDAN achieved the highest classification performance (OA ≈ 0.84). These results indicate that filters that account for local radiometric variability are better able to reduce speckle noise while preserving structural information, which remains important for classification.

This behavior is related to the formation of the SAR signal. The coherent interference of echoes from many small scatterers within a single resolution cell causes speckle. Filtering reduces this noise component, but strong smoothing can also remove small variations in backscatter associated with different scattering processes [87]. Adaptive filters such as Lee Sigma rely on local statistics and reduce speckle noise while preserving edges and intensity gradients. In this way, the filtered image still retains spatial patterns that remain useful for machine learning classification.

In contrast, simple non-adaptive filters based on spatial averaging produced less consistent results. Filters such as Boxcar or Median reduce speckle variance, but their classification results were often lower than those obtained with adaptive filters, as also reported in a previous study comparing non-adaptive and adaptive despeckling approaches [88]. When the filtering window includes surfaces with different scattering responses, the averaging process reduces local backscatter differences between neighboring land-cover types. This effect is particularly visible in heterogeneous landscapes where transitions between surfaces occur within only a few pixels.

The influence of filtering strategies also varied between land-cover types. Homogeneous surfaces, such as water bodies, showed almost saturated classification accuracy regardless of the filtering method used. In this case, the radar signal is dominated by specular surface scattering, resulting in very low backscatter values and limited variability, which simplifies class separability in the SAR feature space [89]. Under these conditions, classification becomes relatively simple in feature space, and filtering plays only a minor role.

Urban environments showed different behavior. Built-up areas typically generate strong double-bounce scattering between vertical structures and the ground surface. Due to this scattering mechanism, backscatter from urban surfaces remains strong and stable even after moderate smoothing, which explains the relatively small differences between filtering strategies in urban classification [90]. As a result, the tested filters produced very similar classification results in the urban AOI, with Lee Sigma, Frost, and Boxcar reaching overall accuracies of 0.88–0.89.

More pronounced differences between filtering strategies appeared in structurally complex environments such as mountainous terrain and cropland areas. In mountainous regions, the SAR signal is influenced by terrain slope, local incidence angle, shadowing, and mixed vegetation cover. In this terrain, adaptive filters such as IDAN reached the highest classification accuracy. Their filtering windows adapt to local radiometric similarity rather than using a fixed spatial kernel. This allows the filter to retain terrain-related backscatter variability while reducing speckle noise.

Agricultural landscapes were also sensitive to the choice of filter. In cropland areas, SAR backscatter changes over short distances because of crop rows, soil roughness, and vegetation cover. Strong smoothing reduces these small variations, making cultivated fields harder to distinguish from nearby surfaces. The performance of the Lee Sigma filter in this environment indicates that selective filtering around the local mean can preserve subtle spatial differences in agricultural SAR signatures.

Forested areas represented the most challenging classification environment. Across all tested configurations, overall accuracy remained considerably lower than in the other AOIs and rarely exceeded approximately 0.56. This result reflects the complex scattering behavior of forest canopies. Vegetation layers contain numerous randomly oriented scatterers such as branches, leaves, and trunks that produce strong volume scattering and multiple reflections. Radar backscatter in forests is therefore highly variable even within the same land-cover class, as widely reported in SAR-based forest studies [90]. Filtering can reduce speckle variance, but it cannot remove variability related to vegetation structure, which makes the separation of forest and surrounding surfaces more difficult.

The experiments showed a clear dependence on polarization. In most AOIs, the VV + VH combination produced higher accuracy than single-channel inputs [91]. The difference was most visible in croplands, where the gain relative to VV reached about 0.10 in overall accuracy. The two channels respond differently. VV is mainly linked to surface scattering and orientation effects, whereas VH captures depolarized returns from vegetation and other complex targets. Using both channels, therefore, provides complementary information that improves the separability of land-cover classes characterized by distinct scattering regimes.

In mountainous areas, dual-polarization input also yielded slightly higher classification accuracy, though the improvement was less pronounced than in cropland AOIs. Radar backscatter in this terrain is influenced by surface roughness, vegetation cover, and local terrain geometry. Using co-polarized and cross-polarized signals provides additional information about these surfaces, allowing better distinction between exposed ground, vegetation patches, and radar-shadow areas, as also reported in studies analyzing SAR backscatter variability under varying terrain conditions [92].

The comparison between ascending and descending acquisitions revealed only moderate differences in classification accuracy. Although the radar viewing geometry affects local incidence angles and backscatter intensity, the relative ranking of filtering strategies remained largely consistent between both orbit directions. Similar findings have been reported in recent Sentinel-1 studies, in which differences between ascending and descending passes influenced backscatter values but had a limited impact on overall classification performance [93].

The experiments evaluating Lee Sigma window sizes showed that stronger smoothing does not necessarily improve classification results. As the window size increased, ENL values increased, and radiometric variability decreased. Also, larger windows smoothed the radar signal and reduced small differences in backscatter that help separate land-cover classes. This effect became noticeable for window sizes larger than 7 × 7. In the analyzed AOIs, the most consistent classification results were obtained with moderate window sizes, particularly between 5 × 5 and 7 × 7, which is consistent with previous studies [81].

The results also show that radiometric image quality metrics alone are not sufficient for evaluating filtering performance in classification tasks. Filters that achieved the highest ENL or the lowest MSE did not always produce the highest classification accuracy, as also demonstrated in recent studies comparing despeckling performance with classification outcomes [94]. Strong smoothing increases statistical homogeneity in the image, but it can also remove spatial patterns that machine learning classifiers use to separate land-cover classes.

Speckle filtering is not limited to radiometric preprocessing. Its effect propagates into the classification stage, where the balance between noise reduction and preservation of spatial structure becomes important. In several cases, stronger smoothing stabilized the signal but reduced local variability, which affected class separability. This means that filter selection cannot be treated independently of the final application. The same configuration may improve image statistics while degrading classification performance. Similar interactions between filtering and classification have been reported in recent SAR processing workflows [95].

The use of DEM thresholds and mask erosion reduced mixing at class boundaries, but at the same time made the classification task somewhat easier than in real-world conditions. Because of this, the reported accuracy may be slightly overestimated, especially for clearly defined classes such as water and mountain areas.

For this reason, radiometric metrics alone are not sufficient for evaluating filter performance. Their interpretation depends on the scene and on how the data are used in subsequent analysis.

This study is based on two Sentinel-1 acquisitions and a limited number of AOIs within a single region. While the results clearly show terrain-dependent behavior, they should not be directly generalized to other seasons, regions, or sensors without further validation.

6. Conclusions

This study evaluates a set of widely used speckle filtering algorithms on Sentinel-1 SAR data, focusing on their effects on image quality metrics and land-cover classification performance. The results differ across scene types. In homogeneous areas such as water or croplands, the choice of filter has only a limited effect, and classification accuracy remains high. In forested and mountainous regions, the response is more sensitive to smoothing. Increasing the filter strength reduces speckle but also removes local details that contribute to class separation. Using both polarizations (VV and VH) generally yields better results than using a single channel. Each channel captures a different part of the backscatter, and their combination changes the separability of classes. The highest accuracies were obtained when dual polarization was used together with adaptive filters. At the same time, stronger filtering does not always improve classification. Better image statistics do not necessarily translate into better classification output. No single filtering strategy performs best in all cases. The choice depends on the scene and on how much structural detail needs to be preserved. Filtering, polarization, and classification, therefore, need to be considered together. Treating them as separate steps can lead to suboptimal results. Future work can extend this analysis to multi-temporal SAR data, multi-class problems, and other regions. Further improvements may come from learning-based despeckling or from combining physical scattering models with data-driven approaches.