Harmonized Landsat-Sentinel 2 Dataset

• HLS is a globally harmonized archive that fuses Landsat and Sentinel-2 imagery into a consistent multi-resolution reflectance product.
• It employs standardized atmospheric corrections, band alignment, and cloud filtering to enable precise time series analyses for vegetation monitoring, biomass estimation, and disaster response.
• The dataset underpins advanced geospatial AI workflows including foundation models, super-resolution, and semantic segmentation, accelerating research in earth sciences and precision agriculture.

The Harmonized Landsat-Sentinel 2 (HLS) Dataset is an operational satellite data product that merges Landsat and Sentinel-2 observations into a temporally dense, radiometrically consistent, and atmospherically corrected global reflectance archive. The HLS dataset provides surface reflectance imagery at 30 m and 10 m spatial resolutions, combining the spectral and temporal strengths of NASA’s Landsat and ESA’s Sentinel-2 missions. This fusion enables robust multi-sensor time series analyses spanning vegetation monitoring, land surface phenology, biomass estimation, precision agriculture, disaster response, and foundational geospatial AI research.

1. Principles of Harmonization and Data Fusion

The harmonization in HLS involves calibrating both spectral and radiometric properties across the different sensors before co-registration at a unified grid. Landsat and Sentinel-2 present differences in their spectral response functions, atmospheric correction workflows, spatial resolutions, and revisit frequencies. In building HLS:

• Atmospheric Correction: Both sensors are processed to surface reflectance using standardized atmospheric corrections, often with Level-2A Sentinel-2 and Landsat Surface Reflectance products. A typical correction for surface reflectance is given by $L_c = (L - L_a)/t$, where $L$ is TOA radiance, $L_a$ is atmospheric path radiance, and $t$ is atmospheric transmittance.
• Band Alignment and Resampling: Sentinel-2’s additional red-edge and SWIR bands at 20 m are bilinearly resampled to match the finer 10 m grid, and similar procedures are used for 30 m products.
• Cloud and Quality Filtering: Scenes with cloud cover $>90\%$ are discarded; additional per-pixel masking is implemented using FMask or similar quality flags.

This process yields a temporally dense, radiometrically stable sequence of multi-spectral images usable for pixel-level time series analysis and direct sensor fusion methodologies (Nachmany et al., 2018, Babcock et al., 2020).

2. Datasets and Label Harmonization Strategies

Multi-sensor harmonized datasets like HLS require reference label sets for downstream tasks such as land cover classification. Common practices and challenges include:

• Reference Label Regridding: Global land cover products such as GlobeLand30 (30 m resolution) are regridded using nearest-neighbor interpolation to align with the target imagery grid (10 m or 30 m, depending on HLS product).
• Temporal Filtering and Agreement Validation: Since reference labels often lag behind sensor acquisitions, temporal mismatches are filtered by comparing regridded labels to classifications generated directly from the imagery, retaining only pixels or scenes where both agree.
• Semantic Harmonization: Cross-regional studies must harmonize label sets, either through grouping detailed classes into standardized higher-level classes or validating field samples in both NALCMS and CORINE schemes (Sokolov et al., 2022).

These approaches reduce label imprecision, mitigate discrepancies arising from spatial and temporal incongruence, and support robust training of machine learning models for semantic segmentation.

3. Foundation Models and Downstream AI Workflows

HLS underpins the development of generalist foundation models in geospatial AI by providing massive volumes of multi-temporal, multi-spectral imagery. A preeminent example is the Prithvi model, which is pre-trained on >1TB HLS data using a self-supervised Masked Autoencoder (MAE) approach—with 3D patch embeddings over space and time—followed by efficient fine-tuning for various earth science tasks (Jakubik et al., 2023):

• Cloud Imputation: Prithvi outperforms conditional GANs in multi-temporal cloud gap imputation, improving SSIM by up to 5.7%.
• Flood, Fire, and Crop Segmentation: Fine-tuned models accurately delineate features for rapid disaster response and agricultural monitoring.
• Data Efficiency: Pre-training enables effective transfer for tasks with <10% of typical labeled data, preserving high IOU, MAE, and SSIM.

A typical masked MAE reconstruction loss is:

$$\text{Loss} = \frac{1}{N} \sum_{p \in \text{masked}} ||x_p - \hat{x}_p||^2,$$

where $x_p$ and $\hat{x}_p$ are the true and reconstructed pixel values, and $N$ is the number of masked tokens.

The resulting workflows and weights are open-sourced and can be directly reused or fine-tuned via standard libraries (e.g., PyTorch, mmsegmentation), facilitating broad adoption in the earth sciences community.

4. Super-Resolution and Multi-Modal Constraints

To overcome native spatial resolution limits of HLS (30 m), recent work adapts generative priors and multi-modal guidance for single-image super-resolution:

• Diffusion Models: DiffFuSR employs a DDPM-based pipeline for 2.5 m super-resolution of Sentinel-2 bands, using a two-stage process: first diffusion model enhancement of RGB bands, followed by fusion-based upscaling of multispectral bands, greatly improving reflectance fidelity and spectral consistency (Sarmad et al., 13 Jun 2025).
• Auxiliary Constraints: LSSR introduces cross-modal attention mechanisms integrating Digital Elevation Model, land cover, temporal metadata, and SAR guidance into Stable Diffusion, with tailored losses (Fourier and NDVI) for spectral fidelity in crop mapping (Yang et al., 27 Oct 2025). NDVI loss is

$$L_{\text{ndvi}} = ||\mathrm{NDVI}(\hat{x}_H) - \mathrm{NDVI}(x_H)||_2^2,$$

enforcing physically meaningful vegetation reconstruction.

• Alignment and Real Sensor Data: Feature Distribution Matching and Histogram Matching are employed for spectral/statistical alignment between heterogeneous sources (Landsat, Sentinel), improving quality before refinement via SwinIR and SR3 diffusion (Shin et al., 30 Jul 2025).

These unified frameworks yield robust spatial and spectral detail, translating into improved object delineation (F1: 0.86 for crop mapping on super-resolved HLS).

5. Temporal, Phenological, and Ecological Applications

Dense HLS time series permit robust estimation of phenological and ecological parameters:

• Phenology: Bayesian hierarchical models fit double logistic curves to vegetation index time series (e.g., EVI), estimating parameters such as seasonal amplitude ($\alpha_2$), dormancy ($\alpha_1$), and transition timing ($\alpha_4, \alpha_7$). Full posterior distributions are computed with Normal, Truncated Normal, or Beta likelihoods as appropriate (Babcock et al., 2020).
• Biomass Change: HLS-derived spectral measures serve as auxiliary variables in model-assisted estimators, dramatically improving the precision of annual/bi-temporal biomass change compared to field plots alone (up to 3× reduction in uncertainty) (Puliti et al., 2020).
• Canopy Height Mapping: U-Net architectures ingest harmonized and normalized Sentinel-2/Sentinel-1 bands to predict GEDI-derived tree heights, achieving MAE ≈ 2.02 m across large forest spans (Schwartz et al., 2022), with error decomposition for diagnostic insight (MSD = SB + SDSD + LCS).

6. Agriculture, Disaster Response, and Operational Deployment

HLS underlies advanced monitoring and policy-critical applications:

• Agricultural Monitoring: Machine learning models trained on radiative transfer simulations (e.g., SCOPE) with HLS imagery estimate crop GPP, outperforming traditional LUE approaches while supporting large-scale, cloud-based deployment (GEE) (Wolanin et al., 2020). Multi-modal fusion with ground-street views and SAR/RADAR data further enhance grassland and crop classification (Choumos et al., 2022).
• Disaster Management: Fused optical (HLS, Sentinel-2/Landsat-9) and SAR (Sentinel-1) maps processed via NDWI and change detection accurately identify flood extents, providing quantifiable overlays with road and population data to inform policy (Nazir et al., 2023).
• Locust Prediction: Deep models (Conv3D, ConvLSTM, and Prithvi-LB) fine-tuned on HLS chips predict breeding grounds, attaining F1 of 81.53% and ROC-AUC of 87.69%—with multi-spectral imagery alone sufficient for effective operational warning (Yusuf et al., 11 Mar 2024).

7. Limitations, Challenges, and Future Directions

Adoption and advancement of HLS require careful consideration of several limitations:

• Sensor Diversity: Despite harmonization, residual differences in sensor properties can necessitate additional spectral normalization, compensation functions, and cross-platform domain adaptation (e.g., HRSemI2I style transfer narrows domain gap for improved segmentation (Sokolov et al., 2022)).
• Label and Temporal Uncertainties: Regridding and agreement filtering between legacy reference products and contemporary imagery must be validated to avoid latent mislabeling.
• Downstream Utility: Model evaluation increasingly considers downstream, task-based metrics (crop boundary F1, NDVI MSE, phenology date uncertainty) rather than only raw image similarity.
• Computation and Data Volume: Foundation and super-resolution models require substantial GPU/TPU resources for both pretraining and inference, with scaling strategies (tiling, Zarr storage) needed for operational processing.

Extending transformer architectures, contrastive learning, efficient diffusion adaptations, and multi-modal fusion open ongoing research avenues to further exploit the harmonized, temporally rich HLS product for continental-scale geoscience, precision management, and environmental forecasting.