Spectral Information Enhancement via STGDFM-Based Spatial-Spectral Fusion: Evaluation Using Sentinel-2-Derived Synthetic Data

Abstract

The inherent trade-offs among spatial resolution, spectral richness, and revisit frequency remain a fundamental limitation of current satellite sensor systems. Spatial-spectral fusion has therefore emerged as an effective strategy to integrate complementary information from heterogeneous observations. This study presents a spatial-spectral fusion framework based on a Spatial Time-series Geostatistical Deconvolution/Fusion Model(STGDFM), with particular emphasis on spectral-profile-based reconstruction. To enable controlled quantitative evaluation, synthetic datasets were generated from Sentinel-2 imagery by simulating complementary coarse- and fine-scale observations. Two implementations were compared: (1) a resolution-matrix-only reconstruction and (2) the full STGDFM framework incorporating spectral profile extraction and residual correction. Quantitative assessment was conducted using the correlation coefficient (CC) and root mean square error (RMSE), focusing on red-edge (RE) and shortwave infrared (SWIR) bands that are not directly available in fine-scale inputs. The results show that the full STGDFM framework generally improves reconstruction consistency across most RE and SWIR bands compared with the resolution-matrix-only implementation. The magnitude of improvement is strongly band-dependent,with more pronounced gains observed in narrow RE bands. Additional experiments conducted in two regions with contrasting land-cover characteristics further indicate that the relative benefit of spectral-profile integration can vary depending on scene heterogeneity and spectral variability. Although the evaluation was conducted using synthetic datasets derived from Sentinel-2 imagery, the findings provide insight into the role of spectral dependency in spatial-spectral fusion. Overall, the proposed framework highlights the importance of jointly reconstructing spatial detail and spectral relationships when extending spectral information in high-resolution Earth observation imagery.

1. Introduction

Recent advances in Earth observation have led to a rapid increase in the availability of satellite imagery with diverse spatial, spectral, and temporal characteristics (Amadin et al., 2025; Kim et al., 2024; Park et al., 2025; Ye et al., 2025). However, no single satellite sensor can simultaneously provide high spatial resolution, rich spectral information, and frequent revisit capability due to inherent trade-offs in sensor design and operational constraints. As a result, satellite image fusion has emerged as an important approach for overcoming these limitations by integrating complementary information from multiple observations (Bai et al., 2026; Ghamisi et al., 2019; Guo et al., 2025; Kim et al., 2020; Shen et al., 2016).

Even with ongoing technological advances toward sensors that offer high spatial resolution, rich spectral coverage, and frequent revisit times, such systems are often associated with substantial costs, operational constraints, and limited data accessibility. For example, a commercial very-high-resolution satellite such as WorldView-3 offers sub-meter spatial resolution with extended multispectral or shortwave-infrared (SWIR) capabilities, yet its data are costly and subject to restrictive licensing policies (European Space Agency, 2026). Similarly, emerging high-revisit commercial constellations, including PlanetScope and BlackSky Gen-3, provide frequent observations but remain limited in spectral diversity and are not always readily accessible for large-scale or long-term scientific applications (BlackSky, 2026; Planet, 2026).

These practical limitations highlight the persistent gap between the growing demand for high-quality satellite-derived information and the constraints imposed by individual sensor systems. Consequently, approaches that can effectively leverage complementary observations from multiple data sources have become increasingly important for reliable and scalable Earth observation applications. Rather than representing a transient methodological trend, data fusion has become a key component of modern remote sensing analysis, supporting a wide range of applications including environmental monitoring, disaster management, and agricultural assessment (Allu & Masapam, 2025; Kim, 2022; Park et al., 2024; Samadzadegan et al., 2025).

Among the various data fusion strategies explored in remote sensing, satellite image fusion is commonly discussed in terms of the complementary trade-offs inherent in sensor design. From this perspective, fusion approaches can be broadly categorized into (1) spatio-temporal data fusion and (2) spatial-spectral data fusion. Spatio-temporal data fusion aims to integrate high-spatial-resolution imagery with high-temporal-resolution observations to generate datasets that are both spatially detailed and temporally continuous (Ghamisi et al., 2019; Swain et al., 2023).

In contrast, spatial-spectral data fusion combines fine-spatial-resolution but spectrally limited imagery (FSLS) with coarse-spatial-resolution yet spectrally rich imagery (CSRS) to produce high-resolution datasets that preserve detailed spectral information (Loncan et al., 2015; Yokoya et al., 2017). These two fusion paradigms address distinct but fundamental limitations of existing satellite systems and form the methodological foundation of many contemporary Earth observation applications.

While spatio-temporal data fusion has been extensively investigated, leading to the development of a wide range of methodologies under diverse assumptions and application contexts (Gao et al., 2006; Kim et al, 2020; Sun & Xiao, 2022; Zhu et al., 2010; Zhu et al., 2016), spatial-spectral data fusion has received comparatively less systematic attention. In many practical settings, spatial-spectral fusion challenges arise from intra-sensor resolution inconsistencies, as observed in sensors such as Moderate Resolution Imaging Spectroradiometer (MODIS) and Sentinel-2, where certain spectral bands are acquired at coarser spatial resolutions than visible channels. Accordingly, numerous studies have focused on spatially enhancing these lower-resolution bands through super-resolution or spatial detail-transfer strategies (Lanaras et al., 2018; Mukherjee & Liu, 2021).

Beyond such intra-sensor resolution harmonization, spatial–spectral fusion has also been widely studied in the context of hyperspectral pansharpening, where spatial information from FSLS data is used to enhance CSRS data (Ma et al., 2024; Qin et al., 2023; Yokoya et al., 2017). In these approaches, the primary objective is typically to improve the spatial detail of spectral channels available only in coarse-scale data. Consequently, the reconstruction is often performed on a band-by-band basis, focusing on spatial refinement of individual spectral channels. While such methods have proven effective for spatial resolution enhancement, they generally place less emphasis on explicitly exploiting the spectral dependency structure across multiple bands within coarse-scale observations.

In contrast, this study explores a spatial–spectral fusion strategy that integrates spectral profile information derived from the full set of spectral channels in the coarse-scale data into the reconstruction process. Rather than treating each spectral band as an independent sharpening target, the proposed framework leverages the spectral relationships among bands to guide the spatial–spectral reconstruction.

To address these limitations, this study explored a spatial-spectral data fusion approach that moves beyond simple band-wise spatial-resolution enhancement and explicitly considers spectral relationships across the dataset. Specifically, a Spatial Time-series Geostatistical Deconvolution/Fusion Model (STGDFM) was employed to link spectral profile information derived from CSRS data with the spatial structures observed in FSLS data (Kim et al., 2020). Originally developed for spatio-temporal data fusion, STGDFM provides a flexible framework for integrating relative spectral relationships. In this context, spectral profiles derived from coarse-scale observations are used to represent the underlying spectral behavior of land-cover types, and these profiles are subsequently integrated with spatial information from fine-scale data through a resolution-matrix-based reconstruction framework.

In this context, this study evaluates the extent to which incorporating spectral profile information provides additional predictive value beyond resolution-based reconstruction alone. While many spatial-spectral fusion methods have primarily focused on enhancing the spatial resolution of individual spectral bands (Ma et al., 2024; Mukherjee & Liu, 2021; Qin et al., 2023), the potential role of explicitly modeling spectral relationships across bands has not been systematically quantified. To address this, the contribution of spectral profile integration is isolated and assessed within a controlled experimental framework.

Building upon this objective, the present study designs a controlled experimental framework to evaluate the feasibility of applying STGDFM to a spatial-spectral fusion problem. Synthetic CSRS and FSLS datasets derived from Sentinel-2 imagery were constructed to enable quantitative assessment of reconstruction accuracy. This design allows the contribution of individual processing components within STGDFM, particularly the role of spectral profile integration, to be systematically examined using reference data. Both qualitative and quantitative evaluations were conducted to analyze the behavior and performance of the proposed framework. While the experiments were limited to Sentinel-2-based synthetic datasets, the analysis provides a controlled proof-of-concept assessment of the spectral-profile-driven fusion strategy and offers insights that may inform future extensions of the framework to real multi-sensor environments.

2. Materials and Methods

The overall flowchart of this study is shown in Fig. 1. In this study, synthetic datasets were employed to evaluate the feasibility of applying the STGDFM framework to the spatial-spectral data fusion problem under a controlled experimental setting. Specifically, the spatial resolution of the original Sentinel-2 image was degraded to generate CSRS data, while only selected channels (i.e., visible and near-infrared [NIR] channels) from the original Sentinel-2 image were used as FSLS data. The original Sentinel-2 image served as reference data for quantitative evaluation. This experimental design enables controlled, repeatable, and pixel-wise quantitative assessment of fusion performance.

Fig. 1. Overview of the proposed STGDFM-based spatial-spectral data fusion approach and the synthetic data evaluation strategy. FSLS data denotes the fine-spatial-resolution but spectrally limited data, whereas CSRS data denotes the coarse-spatial-resolution but spectrally rich data.

The purpose of this study is to evaluate the role of the STGDFM framework in spatial-spectral data fusion by incorporating spectral information derived from spectrally rich CSRS data. In particular, the study focuses on analyzing the influence of spectral profile information extracted from CSRS data on reconstruction performance within the fusion process (Fig. 1). Through this analysis, the study provides a controlled proof-of-concept evaluation of the spectral-profile-driven fusion strategy.

2.1. Study Area and Data

In this study, the applicability of STGDFM was assessed using synthetic datasets constructed to enable a quantitative validation of spatial-spectral data fusion performance. The synthetic data were generated from a Sentinel-2 multispectral image and subsequently employed as the primary experiment datasets. Specifically, two Sentinel-2 images acquired over Illinois, USA, were used for the experiments (Fig. 2). The first dataset (Region A) was derived from an image acquired on 19 April 2018, while the second dataset (Region B) was obtained from an image acquired on 25 July 2018. In both cases, a 2 × 2 km subset was extracted from the original Sentinel-2 scene for the experimental analysis.

Fig. 2. Geographic location of the study areas in Illinois, USA. Insets show Sentinel-2 true-color composite images of the two experimental regions (Regions A and B).

Region A was selected to include diverse land cover types, such as urban areas, water bodies, forests, and agricultural fields, enabling the evaluation of the fusion method under heterogeneous landscape conditions. Region B was additionally selected to represent an area dominated by agricultural land cover, allowing the evaluation of the fusion method under typical cropland conditions. The use of two regions with different land-cover compositions allows a preliminary assessment of whether the observed fusion behavior is consistent across different spatial contexts.

Sentinel-2 provides 13 spectral channels spanning the visible, red-edge (RE), NIR, and SWIR wavelengths (Earthdata, 2025). Among these 13 channels, nine channels – excluding those with a native spatial resolution of 60 m – were used in this study. The CSRS dataset employed all nine channels, with their spatial resolutions uniformly resampled to 200 m. The FSLS dataset utilized four channels corresponding to the visible and NIR wavelengths, retaining their original spatial resolution of 20 m. The specifications of the datasets used in the experiments for Regions A and B are summarized in Table 1.

Table 1. Specifications of the CSRS and FSLS datasets used for the experiments in Regions A and B

This dataset configuration reflects the typical characteristics of existing optical satellite sensors: medium- and low-resolution sensors generally offer broad spectral coverage with high spectral fidelity but relatively coarse spatial resolution, whereas high-resolution satellite sensors provide fine spatial detail with a more restricted set of spectral channels (e.g., KOMPSAT-3/3A, Pleiades Neo, and the NEONSAT series). By emulating these complementary sensor properties, the synthetic datasets provide a controlled experimental environment in which to assess how effectively the proposed spatial-spectral data fusion approach can overcome real-world observational constraints.

2.2. Spatial-Spectral Data Fusion Framework Based on STGDFM

The STGDFM was originally developed for spatio-temporal data fusion, comprising three sequential stages: (1) estimation of low-resolution temporal trends, (2) derivation of high-resolution trend component, (3) residual correction (Kim et al., 2020). In this study, the conceptual structure of STGDFM was adapted to investigate its applicability to spatial-spectral data fusion problems. Particularly, the temporal components of the original framework were reinterpreted in the spectral domain, allowing the model to incorporate spectral profile information derived from spectrally rich datasets.

First, representative spectral profiles for each land-cover type were extracted from the CSRS data. This procedure corresponds to the coarse-scale temporal trends extraction step in the original STGDFM. In the spatial-spectral context, these land-cover-specific spectral profiles serve as coarse-scale descriptors that characterize the spectral behavior of the scene and provide prior spectral information for the fusion process.

Second, a resolution matrix was constructed using co-registered CSRS and FSLS data acquired for the same spectral channels. By applying this matrix, the coarse spectral profiles were deconvolved to the spatial resolution of FSLS data, enabling the estimation of the fine-scale spectral trend component. This step represents the core spatial enhancement process within the spatial-spectral data fusion framework, linking the spectral information of CSRS data with the spatial structures present in FSLS data.

Finally, a residual correction was performed to compensate for the discrepancies not captured by the fine-scale trend estimation. This procedure mitigates potential spectral distortion and reduces spatial discontinuities that may arise during fusion, thereby improving the reconstructed spectral information.

Overall, the adapted STGDFM framework preserves the model’s fundamental estimation-deconvolution-correction logic while reformulating its components to operate in the spectral domain. Through this conceptual transfer, the framework provides a mechanism to integrate spectral profile information with spatial structures into a unified reconstruction process. In this study, the framework is used as a basis to examine the role of spectral profile integration in a controlled experimental setting.

2.3. Experimental Design

To analyze the stage-wise contributions of STGDFM, with particular emphasis on assessing the impact of spectral profile integration on fusion accuracy, two types of outputs were compared in this study (Fig. 1). The first output, referred to as the Full STGDFM Result, was generated by applying the complete STGDFM workflow, including spectral profile extraction, resolution matrix-based spatial enhancement, and residual correction. The second output, referred to as the Resolution-Matrix-Only Result, was produced by applying only the resolution matrix to enhance spatial resolution, without incorporating spectral profile integration or residual correction. This comparison allows the contribution of spectral-profile-based information integration to be explicitly evaluated relative to the resolution-matrix-only configuration. This design is intended to isolate the contribution of spectral profile integration rather than to provide a comprehensive benchmark against existing fusion methods.

The experiments were conducted in both Region A and Region B, as described in Section 2.1, to examine whether the observed fusion behavior is consistent across different spatial contexts and land-cover compositions. For each region, the same experimental protocol was applied to ensure a consistent comparison of fusion performance.

The performance of spatial-spectral fusion was evaluated through comparison with the original Sentinel-2 image used as reference data. A quantitative accuracy assessment was conducted on a per-band basis using the correlation coefficient (CC) and the root mean square error (RMSE). The CC was used to evaluate the degree of spectral similarity and consistency between the fused results and the reference data, independent of absolute radiometric magnitude. RMSE, in contrast, provides a measure of absolute radiometric deviation, capturing both systematic bias and random errors introduced during the fusion process. The combined use of CC and RMSE, therefore, enables a comprehensive evaluation of spatial-spectral fusion performance by jointly assessing relative spectral fidelity and absolute radiometric accuracy within the controlled synthetic experimental framework adopted in this study.

In addition to the quantitative evaluation, a qualitative comparison was first conducted through visual comparison with the reference Sentinel-2 image to examine the spatial continuity and spectral consistency of the fused results. The visual inspection allows intuitive identification of spatial artifacts, edge discontinuities, and band-dependent spectral distortion that may not be fully captured by numerical metrics alone.

An analytical evaluation was also performed to examine the underlying behavior of the spatial-spectral data fusion process. Specifically, a contribution analysis was conducted to investigate the respective roles of individual components within STGDFM by comparing the full model with simplified configurations. This analysis aims to identify how different processing stages influence fusion performance across spectral bands. In addition, a spectral consistency analysis was carried out to assess the extent to which spectral profiles and inter-band relationships were preserved in the fused results. Together, these analytical evaluations provide deeper insight into the mechanisms of spectral information integration and complement the quantitative accuracy assessment.

3. Results

3.1. Qualitative Comparison

To qualitatively evaluate the performance of the proposed STGDFM-based spatial-spectral fusion approach, visual comparisons were conducted between the fusion results and the corresponding reference images. The qualitative assessment focuses on STGDFM’s ability to preserve spatial details while reconstructing spectrally consistent reflectance patterns across different wavelength bands.

Fig. 3 presents a qualitative comparison based on false color composite images, which provide a comprehensive visualization of both spatial structures and spectral relationships across multiple bands for Regions A and B. The comparison includes the Resolution-Matrix-Only Result, the Full STGDFM Result, and the reference Sentinel-2 image.

Fig. 3. Qualitative comparison of false color composites (R: NIR, G: SWIR2, B: RE3) for Region A (top row) and Region B (bottom row): (a) the reference Sentinel-2 image, (b) the full STGDFM result, and (c) the resolution-matrix-only result.

For Region A, the Resolution-Matrix-Only Result exhibits noticeably stronger red tones compared with the magenta patterns observed in the reference Sentinel-2 image. This color shift indicates a degradation in the prediction performance of the RE band, which is assigned to the blue channel in the false color composite representation. As a result, the spectral balance across channels was not properly preserved when only the resolution matrix was applied. In contrast, the Full STGDFM Result more closely resembled the magenta-green tonal balance observed in the reference image, suggesting that the incorporation of spectral profile information may help preserve the inter-band spectral relationships during the fusion process.

For Region B, which was predominantly characterized by agricultural land cover, the visual difference between the two fusion results was less pronounced compared with Region A (Fig. 3). In particular, areas corresponding to bare or sparsely vegetated surfaces were represented with green tones, while cropland areas containing active vegetation appeared predominantly in magenta or purple tones in the false color composite. These spectral patterns were reproduced similarly in both fusion results, indicating that the overall spatial distribution of vegetation-related spectral characteristics was reasonably preserved even when only the resolution matrix was applied.

To further illustrate the band-specific reconstruction characteristics observed in Region A, Fig. 4 presents an enlarged comparison of two selected subareas for the RE2 channel. In the reference Sentinel-2 image, several localized regions exhibit relatively high reflectance values, as highlighted by the yellow boxes. These high-value patterns were not adequately reproduced in the Resolution-Matrix-Only Result, where the corresponding areas appeared attenuated and lacked the contrast observed in the reference image. This indicates that applying only the resolution matrix is insufficient to fully reconstruct the localized spectral variability of the RE2 channel.

Fig. 4. Enlarged comparison of the RE2 band for selected subareas in Region A. The left panel shows the reference Sentinel-2 image with two highlighted regions (Subarea 1 and Subarea 2). The right panels present magnified views of each subarea for (a) the reference Sentinel-2 image, (b) the full STGDFM result, and (c) the resolution-matrix-only result.

In contrast, the Full STGDFM Result reproduced these high-value regions more clearly than the Resolution-Matrix-Only Result, showing spatial patterns and intensity distributions that were more consistent with the reference image. This observation suggests that incorporating spectral profile information through the STGDFM framework may contribute to preserving band-specific spectral characteristics, with more noticeable effects in spectrally sensitive channels such as the RE bands.

In Region A, the Full STGDFM Result more clearly reproduced localized spectral patterns and band-specific reflectance variations observed in the reference Sentinel-2 image, indicating improved reconstruction of RE information compared with the Resolution-Matrix-Only Result. In Region B, where agricultural land cover dominated and spatial patterns were relatively homogeneous, the visual differences between the two fusion results were less pronounced, and both results reproduced the major spectral patterns of the reference image. Nevertheless, the comparisons suggest that the proposed STGDFM-based spatial–spectral fusion framework can produce spatially detailed outputs while preserving overall consistency in spectral reflectance patterns across different spatial contexts. These qualitative observations provide visual evidence of the fusion behavior of the proposed approach and serve as a complementary assessment to the quantitative evaluation presented in the subsequent section.

3.2. Quantitative Evaluation

Following the qualitative comparison, a quantitative evaluation was conducted to objectively assess the reconstruction accuracy of the proposed STGDFM-based spatial-spectral fusion approach. Both the Resolution-Matrix-Only Result and the Full STGDFM Result were compared with the reference Sentinel-2 image using the CC and RMSE (Table 2). The evaluation was performed for both Regions A and B to examine the consistency of fusion performance across different spatial contexts.

Table 2. Band-wise quantitative comparison of the Full STGDFM and the resolution-matrix-only results against the reference Sentinel-2 image using the correlation coefficient (CC) and root mean square error (RMSE)

For Region A, the quantitative difference between the Full STGDFM Result and the Resolution-Matrix-Only Result varied by spectral channel. For RE1, both approaches exhibited nearly identical performance, with CC values of 0.983 (Full STGDFM Result) and 0.982 (Resolution-Matrix-Only Result), and RMSE values of 0.008 and 0.011, respectively. In RE2, the Full STGDFM Result showed improved agreement with the reference Sentinel-2 image, achieving a CC of 0.890 compared to 0.847 for the Resolution-Matrix-Only Result, while the RMSE decreased from 0.032 to 0.023.

More pronounced differences were observed in RE3 and RE4. For RE3, the CC increased from 0.806 (Resolution-Matrix-Only Result) to 0.865 (Full STGDFM Result), and the RMSE decreased from 0.040 to 0.029. Similarly, in RE4, the CC improved from 0.789 to 0.858, while the RMSE decreased from 0.045 to 0.032. Overall, the quantitative improvements were minimal in RE1 but became progressively larger toward RE3 and RE4.

In the SWIR bands of Region A, different quantitative tendencies were observed between SWIR1 and SWIR2. For SWIR1, the Full STGDFM Result showed improved performance compared to the Resolution-Matrix-Only Result. The CC increased from 0.921 to 0.937, while the RMSE decreased from 0.031 to 0.024. In contrast, for SWIR2, the CC value of the Full STGDFM Result (0.916) was lower than that of the Resolution-Matrix-Only Result (0.928). However, the RMSE slightly decreased from 0.023 to 0.021. These results indicate that the quantitative improvements were consistent in SWIR1 for both evaluation metrics, whereas in SWIR2, the two metrics exhibited different tendencies.

For Region B, which is predominantly characterized by agricultural land cover, the quantitative comparison showed generally similar performance tendencies between the two fusion configurations, although the magnitude of the differences varied by band. In the RE bands, the Full STGDFM Result consistently achieved higher CC values and lower RMSE values than the Resolution-Matrix-Only Result. For example, in RE2, the CC increased from 0.747 to 0.797, and the RMSE decreased from 0.028 to 0.025, while in RE3, the CC increased from 0.623 to 0.718, and the RMSE decreased from 0.056 to 0.048. Similar improvements were observed in RE4, where the CC increased from 0.619 to 0.726, and the RMSE decreased from 0.059 to 0.050.

In the SWIR bands of Region B, the two approaches exhibited comparable performance with slight differences depending on the evaluation metric. For SWIR1, the Resolution-Matrix-Only Result achieved a slightly higher CC (0.870) than the Full STGDFM Result (0.835), although the RMSE of the Full STGDFM Result remained comparable. In SWIR2, the Full STGDFM Result showed a slightly higher CC (0.904) and lower RMSE (0.013) compared to the Resolution-Matrix-Only Result (CC – 0.886, RMSE – 0.014).

Overall, the quantitative evaluation across both regions indicated that the STGDFM-based spatial-spectral fusion framework produced reconstruction results generally comparable to those obtained with the Resolution-Matrix-Only approach, with additional improvements in specific spectral bands. In particular, more noticeable improvements were observed in the RE bands, which are spectrally sensitive channels that are difficult to reconstruct using spatial transformation alone. These results suggest that incorporating spectral profile information and residual correction within the STGDFM framework can contribute to improved spectral reconstruction performance in addition to spatial resolution enhancement. At the same time, the magnitude of the improvement may vary depending on scene characteristics and land-cover composition, as observed in the comparison between Regions A and B.

3.3. Spectral Correlation Analysis

A spectral correlation analysis was conducted to examine the band-to-band dependency structure of the fusion results in the RE1-RE4 and SWIR1-SWIR2 regions. This analysis was performed using Region A as a representative dataset in order to investigate how the spectral relationships between bands were preserved during the spatial-spectral fusion process. Visible and NIR bands were excluded from this analysis because they were directly available in the FSLS data and were not subject to spectral reconstruction. The NIR band was used as the reference channel, and correlation coefficients between NIR and each reconstructed band were calculated for the Sentinel-2 image, the Full STGDFM Result, and the Resolution-Matrix-Only Result (Table 3).

Table 3. Pearson correlation coefficients between the NIR band and each RE (RE1-RE4) and SWIR (SWIR1-SWIR2) band for Region A

In the reference Sentinel-2 image, strong correlations with NIR were observed in RE2-RE4, with values of 0.98 (RE2), 0.99 (RE3), and 0.99 (RE4). RE1 exhibited a lower but still substantial correlation of 0.77. In the Resolution-Matrix-Only Result, the correlation coefficients decreased to 0.70 (RE1), 0.78 (RE2), 0.79 (RE3), and 0.79 (RE4). In the Full STGDFM Result, the corresponding values were 0.72 (RE1), 0.84 (RE2), 0.85 (RE3), and 0.86 (RE4). Across all RE bands, the Full STGDFM Result showed higher NIR-centered correlation coefficients than the Resolution-Matrix-Only Result, suggesting a more consistent preservation of inter-band spectral relationships when spectral profile information is incorporated.

For SWIR1, the Sentinel-2 reference value was 0.61, as shown in Table 3. The Resolution-Matrix-Only Result produced a correlation coefficient of 0.60, while the Full STGDFM Result yielded 0.61. For SWIR2, the Sentinel-2 reference value was 0.46. Both fusion results produced identical correlation coefficients of 0.52. In the SWIR region, the differences between the two fusion approaches were small, and both methods produced correlation values close to, or slightly higher than, those of the reference data.

Overall, the results indicate band-dependent variations in the reconstructed spectral dependency structure across the RE and SWIR regions. In particular, the incorporation of spectral profile information into the Full STGDFM framework was associated with more consistent preservation of spectral relationships in the red-edge bands, while the differences between the two approaches were less pronounced in the SWIR bands.

4. Discussion

4.1. Spectral-Profile-Based Integration and Spectral Dependency

The experimental results suggest that the full STGDFM framework produced reconstructions generally comparable to those from the resolution-matrix-only implementation, with differences observed across spectral bands. While the Resolution-Matrix-Only Result improved spatial detail, it showed limited ability to recover band-specific spectral characteristics. In contrast, the Full STGDFM Result—incorporating spectral profile extraction and residual correction in addition to the resolution matrix—exhibited more coherent band-to-band behavior. These observations suggest that integrating spectral profile information may contribute to a more stable reconstruction of inter-band spectral relationships beyond simple spatial redistribution of reflectance values. Accordingly, spatial–spectral data fusion may be viewed not merely as a resolution-enhancement task but as a joint reconstruction process aimed at preserving both spatial structure and spectral consistency (Ma et al., 2024).

The quantitative results reveal that the performance gain achieved by the full STGDFM framework is strongly band-dependent, particularly in the RE and SWIR regions. Rather than exhibiting uniform improvement across all wavelengths, the contribution of spectral profile integration varies according to the intrinsic spectral characteristics and reconstruction difficulty of each band. The following band-dependent analysis primarily focuses on Region A, for which both quantitative accuracy metrics (Table 2) and spectral correlation analysis (Table 3) are available.

In the RE region of Sentinel-2, the improvement was minimal in RE1 but became progressively more pronounced toward RE3 and RE4. RE1 exhibited nearly identical CC and RMSE values between the two implementations (Table 2), suggesting that the resolution matrix alone was already sufficient to reconstruct the spatial distribution of reflectance in this band. The spectral correlation analysis further supports this interpretation (Table 3). In the reference Sentinel-2 image, RE1 showed a moderate correlation with NIR (0.77), and both fusion approaches reproduced similar dependency levels (0.70 for the Resolution-Matrix-Only Result and 0.72 for the Full STGDFM Result). The relatively small difference between the two implementations indicates that explicit spectral-profile integration provides only limited additional benefit for this band.

The progressively larger improvements observed in RE3 and RE4 indicate that spectral-profile-based reconstruction may become more relevant for spectrally more sensitive bands. In the reference Sentinel-2 image, RE2–RE4 exhibited very strong correlation with NIR (0.98–0.99; Table 3). However, the resolution-matrix-only implementation showed reduced dependency levels (0.78–0.79), indicating that spatial redistribution alone was insufficient to preserve the intrinsic spectral relationship fully. The Full STGDFM Result partially restored these correlations (0.84–0.86), consistent with the improvements in CC and RMSE reported in Table 2. This pattern suggests that spectral-profile integration may contribute to maintaining the dependency structure of narrow RE bands, where subtle vegetation-related spectral variations might be preserved alongside spatial detail.

A different pattern emerged in the SWIR region. In SWIR1, consistent improvement in both CC and RMSE (Table 2) was accompanied by stable NIR-centered correlation values (Table 3). The reference correlation was 0.61, and the Full STGDFM Result reproduced this value, while the Resolution-Matrix-Only Result yielded a slightly lower coefficient (0.60; Table 3). Although the difference is small, the alignment with the reference dependency structure indicates that spectral-profile integration may contribute to preserving both magnitude accuracy and inter-band consistency.

In SWIR2, however, the two fusion approaches produced identical NIR-centered correlation coefficients (0.52), both slightly higher than the reference value of 0.46 (Table 3). This result is consistent with the limited or mixed improvements observed in Table 2. The elevated correlation suggests that both fusion approaches may impose a stronger dependency on NIR than observed in the reference Sentinel-2 image, which may be related to the increased difficulty of reconstructing spectrally distant bands. Because SWIR2 lies further from the visible–NIR range, its reconstruction can be considered a more challenging extrapolation task.

To further examine the behavior of the proposed approach under different scene characteristics, an additional experiment was conducted for Region B. This experiment provides further insight into how scene characteristics influence the role of spectral-profile-based fusion. Compared with Region A, Region B is dominated by agricultural land cover and exhibits relatively homogeneous spatial and spectral patterns. Under such conditions, the quantitative differences between the Full STGDFM Result and the Resolution-Matrix-Only Result were generally smaller. Although the full STGDFM framework tended to produce higher CC values and lower RMSE values in most RE bands, the magnitude of improvement was reduced relative to Region A.

An additional observation is that the CC values obtained for Region B were generally lower than those for Region A, even though the qualitative comparison suggested visually coherent reconstruction patterns. This discrepancy can be attributed to the relatively homogeneous land-cover composition of Region B. Because agricultural landscapes often exhibit more homogeneous canopy structures and more uniform spectral characteristics than highly heterogeneous environments, the statistical variance of the reference data is smaller. Under such conditions, correlation-based metrics become more sensitive to small pixel-level deviations, which may reduce CC values even when the reconstructed spatial patterns remain visually consistent.

These results indicate that the relative benefit of spectral-profile integration may depend partly on scene heterogeneity. In heterogeneous landscapes containing multiple land-cover types, spectral variability tends to be larger, and the preservation of inter-band spectral relationships becomes more challenging, potentially increasing the influence of spectral-profile constraints. In contrast, in relatively homogeneous agricultural environments, the spatial redistribution provided by the resolution matrix alone may reproduce much of the dominant spectral structure, resulting in smaller differences between the two fusion configurations.

4.2. Implications and Future Directions

Although this study relied on synthetic datasets derived from Sentinel-2 images to enable controlled quantitative evaluation, the findings provide useful insights into the potential behavior of spatial–spectral fusion approaches in real multi-sensor satellite applications. In operational environments, a fine-scale image frequently lacks specific bands—such as RE and SWIR—making direct pixel-wise validation against reference data impractical. Under such constraints, spatial-spectral data fusion can provide a practical strategy for enhancing the analytical value of high-spatial-resolution images by reconstructing spectrally enriched representations in a physically consistent manner.

The comparison between Regions A and B also suggests that the effectiveness of spatial–spectral fusion may depend on scene characteristics such as land-cover heterogeneity and spectral variability. In landscapes with higher spectral diversity, spectral-profile integration appears to play a more prominent role in preserving complex inter-band relationships. In contrast, in spectrally homogeneous environments, the relative improvement introduced by spectral-profile constraints may become less pronounced. Although this observation is based on the two experimental scenes examined in this study, the results indicate that spatial–spectral fusion performance may be influenced by both spectral characteristics and scene composition.

The results further suggest that spatial-spectral data fusion may remain relevant despite the emergence of next-generation satellite systems that simultaneously provide high spatial and spectral resolution. Sensors such as Planet Labs’ constellation, high-resolution commercial platforms operated by Maxar Technologies, and advanced multispectral missions such as Sentinel-2 illustrate ongoing progress in sensor capabilities (BlackSky, 2026; European Space Agency, 2026; Planet, 2026). However, such systems are still subject to practical constraints, including revisit frequency, swath width, acquisition cost, licensing restrictions, or data accessibility. Under these conditions, fusion-based approaches provide a flexible framework for integrating heterogeneous satellite resources, particularly in institutional or budget-limited operational contexts (Ghamisi et al., 2019; Shen et al., 2016; Yokoya et al., 2017). For example, the proposed framework could be applied to a sensor pairing scenario where high-spatial-resolution imagery (e.g., optical KOMPSAT series, Neonsat, or PlanetScope) is combined with spectrally rich but coarser-resolution data from Sentinel-2, enabling the reconstruction of red-edge or SWIR information at finer spatial scales, which may be particularly useful in applications requiring both detailed spatial mapping and enhanced spectral discrimination (e.g., vegetation or crop monitoring). In practice, careful consideration of differences in spectral response and band harmonization would be required to ensure consistent spectral relationships across the datasets.

Nevertheless, several limitations of this study should be acknowledged. First, the experimental validation was performed using synthetic datasets, which do not fully represent inter-sensor radiometric inconsistencies, bidirectional reflectance effects, or atmospheric discrepancies commonly encountered in real multi-platform fusion scenarios (Samadzadegan et al., 2025). Consequently, the applicability of the proposed framework to operational multi-sensor environments should be further investigated using real heterogeneous satellite datasets.

Second, the experimental evaluation was confined to two study regions derived from Sentinel-2 scenes and therefore covered a limited range of spectral bands and land-cover types, with a relatively small spatial extent. While the additional experiment conducted for Region B provided insight into the influence of scene characteristics on fusion behavior, broader validation across diverse geographic regions and environmental conditions would be necessary to assess the robustness of the proposed approach comprehensively.

Third, qualitative inspection revealed localized block artifacts in the two fusion results (Fig. 4). Although these artifacts did not substantially alter overall spectral tendencies, they may affect fine-scale structural continuity and visual coherence. Such patterns were likely attributable to the resolution-matrix-based reconstruction step, given that similar artifacts were observed in both fusion configurations.

Future research should address these limitations through systematic extensions. Validation using truly heterogeneous sensor combinations—particularly involving sensors with differing spectral response functions and radiometric characteristics—would provide a more rigorous assessment of operational robustness. Expanding the framework to include additional narrowband or hyperspectral channels would further evaluate performance under increased spectral complexity. Further investigation under varying atmospheric, illumination, and seasonal conditions is also necessary to assess stability across realistic observation environments.

To mitigate the observed block artifacts, future work could explore adaptive or multi-scale spatial regularization strategies that enforce smoother transitions across neighboring patches while preserving spectral fidelity. Finally, integration with spatio-temporal fusion frameworks may enable joint spatial–spectral–temporal reconstruction, thereby broadening applicability to dynamic monitoring tasks. Collectively, these directions will contribute to clarifying the generalizability, scalability, and operational feasibility of spectral-profile-based spatial–spectral fusion in real-world Earth observation systems.

5. Conclusions

This study investigated a spectral-profile-driven spatial–spectral data fusion framework based on STGDFM to address the complementary spatial and spectral limitations of satellite observations. Using synthetic CSRS and FSLS datasets derived from Sentinel-2 imagery, the proposed framework was evaluated under a controlled experimental setting that enables quantitative assessment of reconstruction performance.

The results show that incorporating spectral profile information, together with resolution-matrix-based spatial redistribution and residual correction, generally improves reconstruction consistency across spectral bands compared with configurations that rely solely on spatial resolution enhancement. In particular, the improvements were more evident in spectrally sensitive red-edge bands, while the relative contribution of spectral-profile integration varied across wavelengths. Additional experiments conducted for two regions with different land-cover characteristics further indicate that the effectiveness of spectral-profile constraints can depend on scene heterogeneity and spectral variability.

Although the present study relied on synthetic datasets and a limited number of experimental scenes, the analysis provides meaningful insight into the role of spectral dependency in spatial–spectral data fusion. Further evaluation using real heterogeneous satellite sensors and across a broader range of environmental conditions is needed to more fully assess the applicability of the proposed framework. Overall, the findings highlight the potential of spectral-profile-based approaches as a complementary strategy for integrating spatial and spectral information in multi-sensor Earth observation systems.

Author Contributions

Conceptualization, Data curation, Formal analysis, Investigation, Methodology, Validation, Writing–original draft: Kim Y; Project administration, Supervision: Oh K; Writing–review & editing: All authors.

Conflicts of Interest

No potential conflict of interest relevant to this article was reported.

Funding

This research was funded by the Korea Aerospace Research Institute (Grant No. FR26K00).

Data Availability Statement

The data that support the findings of this study are available from the corresponding author upon reasonable request.