Sentinel-2 Multispectral Imagery Overview

• Sentinel-2 MSI is a high-resolution multispectral dataset from ESA's Sentinel satellites featuring 13 co-registered spectral bands across visible, NIR, and SWIR regions.
• It employs precise preprocessing techniques including geometric and radiometric corrections, atmospheric adjustment, and cloud masking to convert TOA to BOA reflectance.
• Advanced fusion methods such as super-resolution, deep learning, and spectral reconstruction enhance index computations and support diverse environmental monitoring applications.

Sentinel-2 Multispectral Imagery refers to the multi-band optical remote sensing data provided by the Sentinel-2A and Sentinel-2B satellites, launched under the European Space Agency (ESA) Copernicus Program. Featuring 13 spectral bands across visible, near-infrared (NIR), and short-wave infrared (SWIR) regions with varying spatial resolutions (10 m, 20 m, and 60 m), Sentinel-2 MSI delivers global high-frequency image coverage, supporting a vast range of scientific and operational Earth observation applications.

1. Sensor Configuration and Data Structure

Sentinel-2 MSI provides 13 spectral bands, partitioned into three spatial resolution groups:

Resolution	Bands	Central Wavelength (nm)	Bandwidth (nm)
10 m	B2, B3, B4, B8	492, 559, 665, 842	65, 35, 30, 115
20 m	B5, B6, B7, B8A, B11, B12	704, 740, 783, 865, 1610, 2190	15, 15, 20, 20, 90, 180
60 m	B1, B9, B10	443, 945, 1373	20, 20, 30

The spectral design targets both land and coastal water monitoring; 10 m bands support geometric fidelity for fine land features, while 20 m/60 m bands provide sensitivity to vegetation, water content, and atmospheric properties. All bands are co-registered, radiometrically calibrated (12-bit quantisation), and recorded with a swath width ≈290 km and revisit frequency of five days using the two-satellite configuration (Schmitt et al., 2019, O'Neill et al., 2021, Revillion et al., 2024).

2. Data Acquisition, Preprocessing, and Atmospheric Correction

Sentinel-2 L1C products are distributed as granules of Top-Of-Atmosphere (TOA) reflectance. Operational workflows for scientific use involve:

• Geometric and Radiometric Correction: L1C scenes are ortho-rectified using rational polynomial coefficients and global DEMs (SRTM/ASTER). Sen2Cor, the ESA standard, is employed for atmospheric correction, converting TOA to Bottom-Of-Atmosphere (BOA) reflectance. Key Sen2Cor steps:
  1. Aerosol optical thickness (AOT) estimation (AOT_METHOD=<DOS1|6S>).
  2. Rayleigh/aerosol/ozone/water vapor compensation.
  3. Optionally, cirrus/snow/cloud adjacency correction.
  4. Reprojection to UTM/WGS84 and spatial resampling (RESAMPLING=<NEAREST|BILINEAR|CUBIC>).
  5. Output: L2A BOA reflectance, per-band geocoded JPEG2000, auxiliary layers (AOT, water vapor, SCL) (Revillion et al., 2024, Schmitt et al., 2019).

• Cloud Masking: Sen2Cor generates Scene Classification Layer (SCL) and probabilistic cloud maps (MSK_CLDPRB), enabling masking via user-defined thresholds (e.g., default 65%) and morphological operators. Custom masks and integration of third-party algorithms (e.g., Fmask) are possible (Revillion et al., 2024).

• Standard Indices: NDVI, NDWI, EVI, and others are computed on corrected BOA data:
  • $$\text{NDVI} = (B_{\text{NIR}} - B_{\text{Red}}) / (B_{\text{NIR}} + B_{\text{Red}})$$
  • $$\text{NDWI} = (B_{\text{Green}} - B_{\text{NIR}}) / (B_{\text{Green}} + B_{\text{NIR}})$$
  • $$\text{EVI} = 2.5 (B_{\text{NIR}} - B_{\text{Red}}) / (B_{\text{NIR}} + 6 B_{\text{Red}} - 7.5 B_{\text{Blue}} + 1)$$ (Revillion et al., 2024, Schmitt et al., 2019).

3. Super-Resolution and Multispectral Data Fusion

Native Sentinel-2 imagery is heterogeneous in resolution. To achieve uniform high-resolution cubes, extensive research addresses super-resolution:

• Band-Independent Geometry Model: High-resolution (HR) bands (10 m) inform sub-pixel geometry. Low-res (20 m/60 m) bands are unmixed via shared sub-pixel reflectances, optimized jointly in a constrained least-squares framework, with geometric weights $W$ band-independent and sub-pixel values $S$ band-dependent. The model enforces exact radiometric consistency on downsampling:

$$L_{b}[\,\lfloor x/R\rfloor,\lfloor y/R\rfloor\,] = \frac{1}{R^2} \sum_{i=0}^{R-1} \sum_{j=0}^{R-1} H^{r}_{b}[\,R\lfloor x/R\rfloor + i, R\lfloor y/R\rfloor + j]$$

This approach achieves Q-index up to 0.994, ERGAS as low as 2.08, and SAM $<2.5^\circ$ in optimal settings (Brodu, 2016).

• Deep Learning-Based Super-Resolution: CNN-based pipelines (e.g., DSen2/VDSen2) exploit globally sampled training data, feeding HR and LR bands (bilinearly upsampled) into ResNet-style architectures. Networks are trained on synthetically downsampled data (scale-invariance assumption), achieving RMSE reductions of ~50% and SAM $\leq 0.76^\circ$ versus classical baselines, with generalization across climates (Lanaras et al., 2018). Modern variations integrate cluster-aware geometry fusion, self-attention, and back-projection (GINet+: PSNR=42.85dB, SSIM=0.9809, SAM=0.7932$^\circ$) (Pereira-Sánchez et al., 5 Aug 2025).
• Multitemporal–Spectral Fusion: DeepSent fuses multispectral and multitemporal stacks in a recursion-based CNN, enhancing all bands to 3.3 m GSD (e.g., cPSNR=43.68, cSSIM=0.9629, SAM=0.0553). Simultaneous fusion surpasses sequential strategies in both simulated and real-world validation (Tarasiewicz et al., 2023).
• Generative and Diffusion Models: Methods using multi-scale GANs or denoising-diffusion architectures (DiffFuSR) can synthesize super-resolved imagery (e.g., all bands to 2.5 m), combine NAIP/WorldStrat high-res data for training, and outperform classical pansharpening (fusion PSNR=32.34 dB, ERGAS=94.8) (Sarmad et al., 13 Jun 2025, Mohandoss et al., 2020).

4. Spectral Extension and Hyperspectral Reconstruction

The broad bands and sparse sampling of Sentinel-2 limit fine material discrimination. Recent work extends Sentinel-2 MSI to virtual hyperspectral imagery via:

• Spectral–Spatial Duality Modeling: ExplainS2A leverages an ADMM-unfolded deep network for spectral super-resolution, constrained by known spectral response profiles, with spatial detail injected from 10 m bands through an explainable fusion module. The resulting 172-band AVIRIS-equivalent cubes attain PSNR=39.42 dB, SAM=1.45°, and SSIM=0.9876 at sub-second inference per million pixels (Lin et al., 21 Apr 2026).
• Transformer Models for MS→HS: Spectral and spatial-spectral Transformers (pretrained on EnMAP/EMIT) reconstruct hyperspectral signatures from Sentinel-2 inputs. Fine-tuned ViT-MAE variants recover 202 bands with MSE $\sim 3.89 \times 10^{-3}$ and SSIM $$\text{NDWI} = (B_{\text{Green}} - B_{\text{NIR}}) / (B_{\text{Green}} + B_{\text{NIR}})$$0, supporting greenhouse gas monitoring and high-fidelity spectral analytics (Gonzalez et al., 26 Feb 2025).

5. Analytical Indices and Application Workflows

Sentinel-2 imagery is routinely processed for downstream geophysical products:

• Time-Series Extraction and Environmental Monitoring: Automated pipelines such as Sen2Chain orchestrate data download, atmospheric correction, masking, and computation of indices (NDVI, NDWI, EVI) across time, with parallel/distributed backends. Output structures follow /{tile_id}/{product}/{YYYY}/{MM}/{DD}/ convention; indices stored as 16-bit integers ($$\text{NDWI} = (B_{\text{Green}} - B_{\text{NIR}}) / (B_{\text{Green}} + B_{\text{NIR}})$$110,000 scaling) (Revillion et al., 2024).
• Cloud/Shadow Masking: Multiple strategies—thresholded probabilistic masks, scene classification, morphology, and band-wise exclusion—are configurable to minimize contamination in derived indices (Revillion et al., 2024).
• Feature Engineering: For aquatic, forest, and fire applications, multiple band combinations (e.g., ND, DG, 4B-ND, DI, NBR) enhance sensitivity to respective biophysical phenomena (Martínez-Ibarra et al., 10 Oct 2025, Xu et al., 2024).
• Object-Based Analysis: Forest mapping integrates deep learning (UNet, ResUNet, AttentionUNet) with OBIA segmentation and SVM fusion, achieving OA $$\text{NDWI} = (B_{\text{Green}} - B_{\text{NIR}}) / (B_{\text{Green}} + B_{\text{NIR}})$$295.6% and IoU $$\text{NDWI} = (B_{\text{Green}} - B_{\text{NIR}}) / (B_{\text{Green}} + B_{\text{NIR}})$$30.91 (Haque et al., 29 Dec 2025).

6. Applications: Land, Water, Atmosphere, and Heritage

Key research and operational domains exploiting Sentinel-2 MSI include:

• Vegetation Height and Type: CNN-based regressors ingesting all 13 bands, atmospheric correction via Sen2Cor, and patchwise aggregation achieve MAE of 1.7 m (Switzerland) and 4.3 m (Gabon) for country-scale 10 m canopy height maps, validated against LiDAR and photogrammetry (Lang et al., 2019).
• Chlorophyll-a/Water Quality: Chlorophyll-a mapping pipelines combine neural-network-based atmospheric correction (C2RCC-C2X-Complex), computed reflectance features and indices, and ensemble ML models (XGBoost, CatBoost, RF, etc.), confirmed against decade-spanning buoy data ($$\text{NDWI} = (B_{\text{Green}} - B_{\text{NIR}}) / (B_{\text{Green}} + B_{\text{NIR}})$$4 up to 0.89 at surface) (Martínez-Ibarra et al., 10 Oct 2025).
• Wildfire Detection: Sen2Fire assembles multispectral+Sentinel-5P dataset, emphasizing SWIR/NBR/NDVI composites. Band selection substantially affects F1 scores (e.g., SWIR composite: 27.9%; RGB+Sentinel-5P aerosol: 17.4%), highlighting importance of physical band choice and fusion with atmospheric data (Xu et al., 2024, Gargiulo et al., 2019).
• Heritage/Archaeology: Multispectral bands (especially B4/B8 NIR) enable detection of surface and shallow-subsurface archaeological structures; fusion with SAR further augments sensitivity to anthropogenic soil disturbance (O'Neill et al., 2021).

7. Datasets, Resources, and Best Practices

Open-access and curated datasets such as SEN12MS (180,662 georeferenced patches at 10 m GSD, all regions/seasons) underpin robust model development and benchmarking. Best practices include:

• Applying atmospheric correction (e.g., Sen2Cor) before further processing.
• Consistent spatial resampling/resolution prior to multi-band or multi-date fusion.
• Automated cloud quality assessment, balanced labelling, and stratified dataset splits to ensure statistical independence.
• Routine validation against in situ or higher-resolution reference datasets (LiDAR, PlanetScope, WorldView-2, AVIRIS) to calibrate and benchmark physical and application-specific accuracy (Schmitt et al., 2019, Revillion et al., 2024, Tarasiewicz et al., 2023).

The architecture of the Sentinel-2 MSI mission, combined with evolving open-source pipelines and deep-learning paradigms, continues to broaden the scope and precision of remote sensing applications reliant on multispectral imagery.