Satellite images usually force a frustrating choice: sharp detail or frequent updates. This paper builds a new Sentinel-2 and Sentinel-3 benchmark to test whether spatiotemporal fusion can genuinely bridge that gap. Multiple sensors are used to overcome the limitations inherent in individual data, but sensor lifespans are finite.

That creates an ongoing need to explore and integrate new sensors. The response is a new benchmark remote sensing dataset based on Sentinel-2 and Sentinel-3 OLCI, comprising more than six hundred eighty-nine image pairs. The dataset spans wide temporal and spatial variations, capturing diverse landscapes, ecosystems, and weather conditions, and it is publicly accessible.

A comprehensive quantitative and qualitative comparison of widely used spatiotemporal fusion methods is also conducted using this dataset. Dense time-series data are crucial for monitoring crop yields, natural hazards and disasters, and changes in land use. Yet a satellite product with high spatial and temporal resolutions simultaneously is almost impossible to find.

MODIS provides daily reflectance images at spatial resolutions from two hundred fifty meters to one kilometer, while Sentinel-3A and Sentinel-3B capture images at three hundred meters. Landsat, SPOT-5, and Sentinel-2A and Sentinel-2B offer finer resolutions between ten and sixty meters, but their revisiting cycles are longer.

Sentinel-2 revisits every five days, Landsat every sixteen days, and SPOT-5 every twenty-six days. Integrating different sensors can address this trade-off and generate dense time series data. Remote sensing data fusion has advanced significantly, but a thorough examination of available methods using benchmark datasets is still lacking.

Consistent evaluation is difficult because suitable datasets rarely cover varied locations, scenarios, landscapes, and new sensors. Earlier public datasets compared methods such as STARFM and ESTARFM using MODIS and Landsat images, including data from two sites with contrasting spatial and temporal variability.

A separate public dataset for spatiotemporal models consisted of MODIS and Landsat pairs. Existing benchmark datasets rely exclusively on MODIS and Landsat data, making it difficult to understand fusion-model efficacy across different sensors.

They also have limited numbers of paired images and a lack of image diversity, which are insufficient for a fair comparison. Spatiotemporal fusion methods merge data sources with similar spectral ranges to provide time series with high spatial and temporal resolutions.

Typically, one sensor provides high spatial but low temporal resolution, while another provides low spatial but high temporal resolution. Although STF methods differ in prediction accuracy, computing efficiency, and input-data requirements, few studies compare them directly.

There is no comparative study of STF methods using Sentinel-2 and Sentinel-3 OLCI, so fusion methods need testing across different scenarios, sensors, and satellites. The new benchmark dataset is based on Sentinel-2 and Sentinel-3 images and introduces new European sensors into STF algorithms.

It is large-scale in spatial resolution and total paired-image count, and it provides a high spatial-resolution ratio between coarse and fine images. The dataset includes several ecosystems, landscapes, weather conditions, and temporal and spatial variances.

The study compares five commonly used STF methods and evaluates them quantitatively and qualitatively on the benchmarked dataset. The benchmark includes fifteen sites spanning diverse ecosystems and geographic locations, with a total of six hundred eighty-nine Sentinel-2 and Sentinel-3 pairs.

The sites introduce challenges ranging from rapid environmental changes to urban expansion and coastal ecosystem dynamics. Each pair combines Sentinel-3, with sixteen bands at three hundred meters, and Sentinel-2, with ten bands at spatial resolutions of ten, twenty, and sixty meters.

Multiple bands and diverse geographic conditions support rigorous testing, while varied environmental conditions help identify limitations and failure modes in STF techniques. Figure one maps the geographic locations of the sites included in the dataset, with red markers distributed across North and South America, Europe, Africa, Asia, and Australia.

The authors assembled fifteen sites spanning diverse ecosystems and geographic settings, creating a benchmark that includes challenges such as rapid environmental change, urban development, deforestation, and coastal dynamics. This broad spatial coverage matters because it tests spatiotemporal fusion methods across varied landscapes rather than a single region.

Table one catalogs the spectral bands used from Sentinel-two and Sentinel-three, giving each band’s wavelength range and spatial resolution. Sentinel-two includes bands such as Band two at 458 to 523 nanometers and ten meters, while Sentinel-three lists bands from O-a-one through O-a-twenty-one, each at 300 meters.

This reference matters because the benchmark combines these sensors across 689 image pairs, so their different spectral and spatial specifications define the input data for evaluating fusion algorithms. The sites are distributed across North America, South America, Africa, Europe, Asia, and Australia, ensuring global representation.

The dataset covers arid deserts, temperate forests, tropical rainforests, coastal zones, and high-altitude regions, allowing STF models to be evaluated under different environmental conditions. Some locations were selected for susceptibility to rapid transformations, including wildfire-prone landscapes, expanding urban areas, and vulnerable coastal ecosystems.

Cloud-free data availability was critical for ensuring that Sentinel-2 and Sentinel-3 observations were consistently available for each location. The first preprocessing step is atmospheric correction, which ensures the accuracy and consistency of surface reflectance values.

Atmospheric correction removes effects such as scattering and absorption caused by gases, aerosols, and water vapor, standardizing data across acquisition times and sensor platforms. The study uses the Image Correction for Atmospheric Effects algorithm for Sentinel-2 and Sentinel-3 OLCI level data.

A uniform correction process is applied across both sensors, minimizing potential biases and enhancing dataset reliability for fusion applications. ENVI five point six point one co-registers a single Sentinel-2 and Sentinel-3 image pair, and AROSICS then co-registers the entire time series.

AROSICS uses cross-correlation in Fourier space for accurate co-registration, which matters because geolocation misalignments can occur even within one sensor time series. Without proper time-series co-registration, these errors would significantly degrade fusion quality, especially for temporally consistent generation.

Sentinel-3 OLCI images are resampled with the Nearest Neighbor method to match Sentinel-2 spatial resolution, while all spectral bands are preserved. The dataset is organized by sites and date, freely available, and provided in WGS eighty-four slash UTM zone forty-eight north projection with TIFF format.

The benchmark analysis assesses representative STF methods on the dataset. Its objective is a thorough examination of current state-of-the-art performance specifically within the framework of Sentinel imagery. The resulting insights improve research comprehension and lay a foundation for future work in the field.

ESTARFM was proposed to enhance STARFM performance in heterogeneous landscapes and requires two coarse-fine image pairs from prior and posterior dates, plus a coarse image at the prediction date. The two fine images search for similar pixels within a moving window centered on the target pixel.

Spectral similarity determines the weights for similar pixels, while linear regression calculates conversion coefficients for each similar pixel in the search window. The weight function and conversion coefficients are combined to predict the fused image at the prediction date.

FSDAF is a hybrid method that combines weighted-function-based and unmixing-based methods with spatial interpolation. It requires at least a coarse-fine pair on the base date and a coarse image on the prediction date. The fine image generates a classification map, and the coarse images estimate temporal changes for each class between the two dates.

FSDAF makes temporal and spatial predictions, distributes temporal residuals using a weighted function, and estimates the final prediction with neighborhood information. Fit-FC is a linear regression-based method designed for fusion of Sentinel-2 MSI and Sentinel-3 OLCI images.

It fits local linear regression models between coarse images at the base and prediction dates within a moving window, then applies the coefficients to the fine image at the base date. Spatial filtering addresses blocky artifacts, using spectrally similar pixels within the moving window to define a weighted function.

Residual compensation downscales coarse-image residuals with cubic interpolation, and the updated residual is added back to preserve spectral information. SFSDAF is an enhanced FSDAF method that incorporates subpixel class fraction change information and requires one coarse-fine image pair at a prior date.

ISODATA generates a land-cover map from the fine image, while endmembers are calculated as average fine-resolution pixels for each class and soft classification is applied. A coarse-resolution class-fraction change image is calculated between the two dates, then downscaled using bicubic and Thin Plate Spline interpolation.

The downscaled image is added to the fine-resolution class fraction at the first date, and spatial interpolation combines the temporal and spatial predictions. The fusion methods receive a coarse-fine Sentinel-3 and Sentinel-2 pair from the base date, together with a coarse image at the prediction date.

The fine image at the prediction date serves as ground truth for comparing the prediction results. Peak Signal to Noise Ratio is also employed as a quantitative assessment of spatial reconstruction quality and overall bias between the predicted and original images.

A lower root mean square error indicates higher accuracy, while Cross-Correlation has an ideal value of one and evaluates geometric distortion. Structural Similarity quantifies spatial similarity in overall structure, and a higher Structural Similarity value indicates better results.

Peak Signal to Noise Ratio evaluates spatial reconstruction quality and overall bias, with a higher value indicating higher-quality spatial reconstruction. Figure six compares the coarse Sentinel-three images, the higher-resolution Sentinel-two observations, and predictions from five fusion methods across Waterbank, Neom city, and Maspalomas.

The false-color views make spatial changes visible, including irrigation patterns, coastal features, and urban or land-cover details, while the predicted images can be compared directly with the observed Sentinel-two image on the prediction date. The authors use these examples to show that all methods generally capture spatial detail despite the large resolution gap, while some fine-scale changes remain difficult.

Figure seven zooms into the boxed regions from the broader comparison, showing Sentinel-two imagery at the prior and prediction dates alongside outputs from STARFM, ESTARFM, FSDAF, Fit-FC, and SFSDAP. The rows highlight different sites and landscapes, including center-pivot irrigations whose false-color appearance changes from green to white while their size and boundaries remain similar.

This close-up matters because it exposes how difficult it is for spatiotemporal fusion methods to preserve small spatial details and represent sudden or continuous temporal changes in heterogeneous scenes. Cross-Correlation, root mean square error, Structural Similarity, and Peak Signal to Noise Ratio are calculated for three sites to better assess prediction accuracy.

Tables four, five, and six present detailed quantitative assessments for every band used in the comparison of the aforementioned methods. All methods fuse different bands within the bandwidth of Sentinel-2 and Sentinel-3 OLCI, including a higher spatial-resolution ratio between coarse and fine images.

ESTARFM obtains competitive results, with metrics closer to optimal values than the other methods in all six bands. For Neom City band six, ESTARFM improves Structural Similarity from zero point four to zero point eighty-eight and consistently achieves higher Peak Signal to Noise Ratio values across sites and bands.

SFSDAF also shows good results at all three sites, with metrics comparable to ESTARFM. Table 4 reports band-by-band results for STARFM, ESTARFM, FSDAF, Fit-FC, and SFSDAF at the Waterbank site on 08 June 2019. It uses four metrics—CC, RMSE, SSIM, and PSNR—to quantify prediction accuracy across six bands, with the reported values showing substantial variation by metric and band.

The authors include this detailed assessment to evaluate how well these spatiotemporal fusion methods reproduce image information across different sensor bands. Table Five reports quantitative results for the Neom City site on twelve July twenty eighteen across six bands and four metrics: CC, RMSE, SSIM, and PSNR.

The authors compare STARFM, ESTARFM, FSDAF, Fit-FC, and SFSDAF; for example, ESTARFM’s CC values range from zero point nine seven two seven to zero point nine eight six four, while its RMSE values range from zero point zero two four five to zero point zero five seven zero. This matters because the table provides a band-by-band numerical assessment supporting the paper’s discussion of method behavior at this continuously changing site.

Table six reports four metrics across six spectral bands for the Maspalomas site on twenty-eight September twenty nineteen, comparing STARFM, ESTARFM, FSDAF, Fit-FC, and SFSDAF. It shows correlation coefficient, or CC; RMSE; SSIM; and PSNR, making both agreement and image-quality differences visible across bands.

The table matters because it provides a band-by-band quantitative check of the experimental results, complementing the paper’s visual assessments. Figure eight summarizes average RMSE, CC, SSIM, and PSNR across six bands, comparing actual and predicted images for five STF methods at the Waterbank, Neom City, and Maspalomas sites.

The heatmaps preserve variation across dates and locations, including the larger temporal changes associated with particular Neom City pairs. This matters because the visual comparison complements the numerical evaluation, showing how method performance responds to different landscape heterogeneity and temporal variation.

Performance variation in visual and quantitative assessments can be attributed to the number of pairs used in the fusion process. ESTARFM requires two pairs from prior and posterior dates, and this additional information about temporal changes enables more accurate predictions.

Even with multiple pairs, all methods experience challenges in certain pixels, resulting in inaccurate predictions in some cases. The five STF methods show no significant differences in landscapes with low heterogeneity and low temporal variance. Their performance varies in landscapes with higher heterogeneity and greater temporal variance, highlighting sensitivity to those conditions.

ESTARFM is more suitable for heterogeneous landscapes and sudden events because it more accurately captures structural details, even though it may not always have the lowest errors. When coarse-fine image pairs are not always available, SFSDAF or STARFM delivers competitive results while working with a minimum number of image pairs.

The benchmark shows that ESTARFM is especially suitable for heterogeneous landscapes and sudden events, while SFSDAF and STARFM remain competitive when only a minimum number of image pairs is available.