Satellite-Based Land Surface Temperature

• Satellite-based LST is the radiative skin temperature of the Earth measured from space using thermal infrared sensors to capture energy exchange at the surface.
• Key methodologies include the Split-Window algorithm, NDVI-guided emissivity adjustments, and deep learning fusion techniques, enabling high-resolution, robust estimates.
• Applications span urban heat island mapping, water resources management, and climate diagnostics, while addressing uncertainties from sensor calibration and atmospheric effects.

Satellite-based Land Surface Temperature (LST) is defined as the radiative skin temperature of the land’s surface, retrieved from spaceborne sensors, and represents the integral response of the land to energy partitioning at its interface with the atmosphere. LST is a central variable in terrestrial climate, surface energy balance, urban heat island studies, water resources management, and environmental health diagnostics. Satellite sensors, exploiting primarily the thermal infrared (TIR) and sometimes microwave regions, enable consistent global measurement of LST, facilitating cross-disciplinary Earth system studies at scales inaccessible to classical ground sensing.

1. Physical Basis and LST Retrieval Algorithms

Satellite-based LST quantifies the Planckian radiance emitted by the Earth's surface in the TIR (typically 8–14 μm), corrected for atmospheric transmission and surface emissivity ($\epsilon$). The most widely used physical approach for LST retrieval is the Split-Window (SW) algorithm, which leverages two neighboring TIR channels to correct for atmospheric water vapor effects and residual emissivity variability. The general SW form is:

$$T_s = T_i + C_1 (T_i - T_j) + C_2 (T_i - T_j)^2 + C_0 + (C_3 + C_4 w)(1-\epsilon) + (C_5 + C_6 w)\Delta\epsilon$$

where $T_i$, $T_j$ are the brightness temperatures, $w$ is column water vapor, $\epsilon$ the mean emissivity, and $\Delta\epsilon$ the emissivity difference between bands; $C_0$–$C_6$ are coefficients empirically or physically derived, as in modern coupled mechanism/ML frameworks (Xie et al., 5 Sep 2025).

Physically-based radiative transfer (e.g., MODTRAN) simulations generate synthetic datasets to train or constrain such algorithms. More recent approaches embed explicit dependence on land cover-derived emissivities, often using NDVI-guided methods (e.g., NDVI Thresholds Method: $\epsilon = 1.0094 + 0.047\ln(\mathrm{NDVI})$ for NDVI in [0.157,0.727]) (Serban et al., 2011). The adoption of algorithms requiring minimal ancillary data (e.g., Jimenez-Munoz and Sobrino: only total column water vapor) allows operation in data-sparse settings, while still producing robust LST estimates.

2. Preprocessing, Ancillary Variables, and Data Harmonization

LST retrieval relies on radiometric calibration of satellite TIR signals, geometric and atmospheric correction (e.g., Dark Object Subtraction for reflectance), and precise co-registration with auxiliary data. Land cover classification—often via supervised parallelepiped algorithms using multispectral bands—is routinely performed to aid in the parameterization of surface emissivity and support urban/environmental applications (Serban et al., 2011). NDVI and ancillary spectral indices are not only used for emissivity calculation, but (when available at higher resolution) serve as guides for sub-pixel detail injection in downscaling and super-resolution frameworks (Ait-Bachir et al., 3 Feb 2025).

Multi-sensor data harmonization is critical. Modern frameworks employ both temporal normalization (correction for overpass time differences via diurnal cycle models, e.g., GOT09) and sensor normalization (e.g., global linear model crosswalks between Landsat-MODIS LST) to facilitate spatiotemporal fusion (Ma et al., 2021).

3. Reconstruction, Gap-Filling, and Spatio-Temporal Fusion

Satellite LST products are plagued by gaps due to cloud contamination, orbit gaps, and sensor malfunctions. Gap-filling techniques fall into three major categories:

• Spatial approaches: Interpolate missing pixels using neighboring LST, with methods such as inverse distance weighting or kriging.
• Temporal approaches: Rely on temporal dynamics at given locations, using Fourier/harmonic analysis, wavelet transforms, and diurnal cycle models.
• Spatiotemporal fusion: Combine both, as in the general form $$\mathrm{LST}(x_0, y_0, t_0, s_0) = \sum p_{x, y, t, s} f(\mathrm{LST}(x, y, t, s))$$ (Wu et al., 2019).

More advanced methods reconstruct cloudy pixels by integrating passive microwave LST (less susceptible to clouds) or via surface energy balance modeling with local meteorological constraints. Recent hybrid models blend these with data-driven or deep learning frameworks.

Crucially, high spatiotemporal resolution is achieved through spatiotemporal fusion: combining high-frequency, low-spatial (e.g., MODIS 1 km, daily) and low-frequency, high-spatial (e.g., Landsat 30 m, 16-day) products. Methods include weighted function-based models (e.g., STARFM, ESTARFM), unmixing, and hybrid models (Wu et al., 2019, Ma et al., 2021). Deep learning now dominates, with CNNs, autoencoders, GANs, and transformer-based methods trained to learn nonlinear fusion mappings, leading to state-of-the-art results (Bouaziz et al., 21 Dec 2024).

4. Super-Resolution and Downscaling

Achieving LST retrieval at resolutions finer than native sensor capabilities necessitates super-resolution methods. Traditional regression and statistical sharpening methods (ATPRK, DMS, TsHARP) are outperformed by deep neural networks:

• Residual Learning U-Nets: Multi-residual U-Net architectures enhance super-resolution by regressing the residual between bicubic-interpolated coarse LST and the target fine LST, with skip and residual connections maintaining texture and reducing gradient issues (Nguyen et al., 2022).
• Scale-Invariance-Free Approaches: Training at full spatial resolution (rather than relying on scale-invariance across resolutions) and guiding texture synthesis using co-registered, high-resolution NDVI enables better recovery of fine spatial structures. Objective functions combine degradation consistency and texture fidelity, evaluated with perceptual and Fourier domain metrics (Ait-Bachir et al., 3 Feb 2025).
• Modality-Conditional Fusion: Networks using Modality-Conditional Large Selective Kernel (MoCoLSK) mechanisms dynamically adjust receptive fields and fuse multi-modal data (LST, multispectral, DEM) for guided downscaling, with open-source benchmarks supporting standardized comparison (Dai et al., 30 Sep 2024).

These strategies yield substantial gains in RMSE, SSIM, and visual quality, particularly in heterogeneous regions where fine-scale thermal variability encodes distinctive environmental signals.

5. Challenges, Limitations, and Best Practices

Persistent limitations are shaped by data gaps (clouds, revisit cycles), sensor differences, atmospheric effects, and surface heterogeneity. Both classical and deep learning methods propagate retrieval uncertainties from sensor calibration, atmospheric water vapor, and emissivity errors—the last being of lesser sensitivity overall but nontrivial for specific surface types (Xie et al., 5 Sep 2025). Hybrid mechanism-ML frameworks (MM-ML) explicitly constrain ML outputs with physical equations (radiative transfer, split-window), yielding improved accuracy (MAE ~1.84 K, RMSE ~2.55 K), especially under extreme atmospheric regimes (bias reduction >50% vs. pure SW or ML baselines).

Recent approaches feature uncertainty quantification via deep ensembles and Gaussian process post-processing to provide prediction intervals for each LST estimate, critical for risk assessment and public health applications (Liu et al., 20 Feb 2025).

Despite these methodological advances, LST is not a direct proxy for near-surface air temperature (SAT) or human thermal comfort. The radiative surface temperature measured from space can exceed SAT by 20–30 °C under clear-sky conditions, and is sensitive to surface orientation, material, and viewing geometry. Using LST as a surrogate for SAT systematically mischaracterizes urban heat hazard, inflating adaptation benefits and distorting spatial risk patterns. Policy and scientific studies are therefore advised to treat LST as a distinct physical variable—useful for surface energy budgets and climate diagnostics, but not directly for pedestrian thermal exposure— and to clearly label LST-derived findings, integrating complementary in situ and model-based air temperature products to inform adaptation policy (Zhan et al., 20 Sep 2025).

6. Practical Applications and Public Datasets

Satellite-based LST is foundational in:

• Urban heat island mapping: High-resolution, gapless LST time series (via ISLAND, DELAG, WGAST) enable quantification of UHI frequency and intensity across land cover classes, supporting targeted interventions and long-term monitoring (Liu et al., 2023, Liu et al., 20 Feb 2025, Bouaziz et al., 8 Aug 2025).
• Water resources management: Joint retrievals of NDVI, LST, and land cover class allow the assessment of vegetation water stress, body identification, and drought monitoring, informing irrigation planning (Serban et al., 2011).
• Building energy efficiency: Coarse-resolution LST, fused with other remotely sensed imagery (street view, aerial), provides a complementary signal for large-scale energy efficiency classification, although limitations in spatial resolution currently reduce its standalone effectiveness (Mayer et al., 2022).
• Diurnal and seasonal monitoring: Integration with land surface models (e.g., CLM 5.0), deep ensemble reconstructions, and spatiotemporal fusion models generate gapless, high-frequency LST products capable of resolving fine-scale diurnal cycles and seasonal patterns (Ma et al., 2021, Liu et al., 20 Feb 2025).
• Climate and ecosystem studies: Satellite-derived LST supports global assessments of surface energy partitioning, as demonstrated in physics-based analytical models grounded on maximum power principles (Gupta et al., 2022).

Several datasets and toolkits are now public, including pre-computed high-resolution, gapless LST archives for major U.S. cities (ISLAND), benchmarks for high-resolution spatiotemporal fusion and downscaling (GrokLST), and testbeds for super-resolution approaches with ASTER-MODIS pairs, anchoring reproducible research and standardized performance comparisons (Liu et al., 2023, Dai et al., 30 Sep 2024, Ait-Bachir et al., 3 Feb 2025).

7. Recent Trends and Outlook

Deep learning methods, particularly those integrating mechanism-informed priors and physics constraints, are now the frontier of satellite-based LST retrieval, fusion, and super-resolution. Benchmarks indicate that advances in multi-modal feature fusion, dynamic receptive field adjustment, and uncertainty quantification have led to significant improvements in retrieval accuracy, texture fidelity, and robustness to atmospheric and land cover variability (Dai et al., 30 Sep 2024, Xie et al., 5 Sep 2025, Bouaziz et al., 8 Aug 2025). Weakly-supervised generative networks now generate daily LST at 10 m resolution, resolving many long-standing trade-offs between spatial and temporal detail (Bouaziz et al., 8 Aug 2025, Bouaziz et al., 30 Jul 2025).

Ongoing challenges include developing better domain adaptation for heterogeneous global settings, integrating LST with in situ and air temperature products for human-relevant hazard estimation, and addressing the propagation of uncertainty and error through complex fusion architectures. Further, the role of LST in urban adaptation and public health policy is the subject of ongoing debate, emphasizing the need for careful interpretation and multidimensional analysis in applied settings (Zhan et al., 20 Sep 2025).

Researchers are advised to adhere to best practices: (1) explicitly contextualizing the meaning and scope of LST, (2) combining LST with complementary data and models for process-based inference, and (3) employing standardized, open-source toolkits and benchmarks to ensure reproducibility and comparability in algorithm development and evaluation.