Land Surface Temperature Super-Resolution With a Scale-Invariance-Free Neural Approach: Application to MODIS

Romuald Ait-Bachir, Carlos Granero-Belinchón, Aurélie Michel, Julien Michel, Xavier Briottet, Lucas Drumetz

TL;DR

Researchers have developed a new "Scale-Invariance-Free" neural network approach to sharpen Land Surface Temperature (LST) satellite data from 1 km to 250 m resolution. By moving away from the flawed assumption that thermal patterns look the same at all scales, and by using high-resolution vegetation data (NDVI) to guide texture synthesis, their models significantly outperform traditional statistical and deep learning methods in capturing realistic surface details.

What Problem Does This Paper Address?

In satellite remote sensing, there is a frustrating "thermal trade-off." Sensors like MODIS provide data twice a day (high temporal resolution) but only at a coarse 1 km per pixel (low spatial resolution). Conversely, sensors like ASTER provide sharp 90 m resolution but only visit the same spot once every 16 days. To study local phenomena like urban heat islands or forest fires, we need both high frequency and high detail.

Existing "super-resolution" (SR) methods usually rely on a scale-invariance hypothesis. They train a model to turn a 4 km pixel into a 1 km pixel and then assume the same logic applies to turning a 1 km pixel into a 250 m pixel. However, the physical relationship between temperature and land cover changes across scales. This paper tackles this "scale-invariance trap" by proposing a training method that works directly at the target resolution without needing high-resolution temperature labels.

Methods

The authors introduce the Scale-Invariance-Free Convolutional Neural Network for Super-Resolution (SIF-CNN-SR).

The core innovation is a self-supervised training strategy inspired by variational optimization. Instead of comparing the output to a "ground truth" high-res image (which doesn't exist for MODIS LST), the loss function enforces two constraints:

  1. Consistency: If you degrade the high-resolution output back to 1 km, it must match the original coarse observation.
  2. Synthesis: The high-frequency textures in the output should be "informed" by high-resolution NDVI (Normalized Difference Vegetation Index) data, which is available at 250 m.

They implemented two versions:

  • SIF-CNN-SR1: Uses Sobel filters to define and transfer gradients/textures.
  • SIF-CNN-SR2: Uses high-pass filtering for texture definition.

The underlying architecture is a Multi-Residual U-Net, a deep learning structure well-suited for image-to-image tasks.

Datasets

  • Training/Validation: 1,095 pairs of MODIS LST (1 km) and MODIS NDVI (250 m) images over Central Europe (2017–2019), sliced into over 8,000 patches.
  • Evaluation (The Benchmark): A newly released dataset of 79 concomitant ASTER (90 m) and MODIS (1 km) LST images. This is significant because ASTER provides a real-world high-resolution reference to validate the upscaled MODIS data, rather than relying on simulated downsampling.

Key Insights and Ideas

  • Breaking the Scale Barrier: By using the "consistency" property, the model learns to sharpen images without ever seeing a "perfect" high-res temperature map during training. This bypasses the errors introduced by assuming physics works the same at 4 km as it does at 250 m.
  • NDVI as a Texture Proxy: Since temperature is often inversely related to vegetation (plants are cooler), the model "borrows" the sharp edges found in vegetation maps to place sharp edges in the temperature maps.
  • Texture-First Evaluation: The authors argue that standard metrics like RMSE (Root Mean Square Error) are insufficient because they favor "blurry" images that play it safe. They emphasize LPIPS (a perceptual metric) and Fourier space analysis to prove their model actually restores realistic high-frequency textures.

Key Results

  • Superior Texture: SIF-CNN-SR1 achieved the best LPIPS score (0.28), indicating it produces the most visually and structurally realistic textures compared to state-of-the-art methods like ATPRK or DMS.
  • Accuracy: Both SIF models achieved an RMSE of ~1.9 K, outperforming traditional statistical methods (2.1–2.3 K) and a standard scale-invariant CNN.
  • Speed: The neural network approach is incredibly efficient, with an inference time of 0.03 seconds per image—hundreds of times faster than some statistical methods like ATPRK (19.56 s).
  • Spectral Fidelity: Fourier analysis showed that SIF-CNN-SR1's attenuation spectrum closely follows the "real" spectrum of ASTER data, whereas other methods either over-sharpen or under-sharpen the results.

Limitations and Future Work

  • The NDVI Assumption: The model assumes a strong relationship between vegetation and temperature. This holds for most land, but fails for water bodies or during winter when the LST-NDVI correlation weakens or flips.
  • Sensor Bias: There is an inherent radiometric bias between MODIS and ASTER sensors (often 0–2 K), which complicates the evaluation.
  • Nighttime Data: The current model only uses daytime data (since NDVI requires sunlight). Future work is needed to address nighttime super-resolution where NDVI isn't available.
  • Higher Resolutions: The authors aim to test the model on upcoming missions like TRISHNA or LSTM, which will provide even finer resolutions (around 50 m).

Summary updated

BibTeX
@article{aitbachir2025land,
  author = {Romuald Ait-Bachir and Carlos Granero-Belinchón and Aurélie Michel and Julien Michel and Xavier Briottet and Lucas Drumetz},
  title = {{Land Surface Temperature Super-Resolution With a Scale-Invariance-Free Neural Approach: Application to MODIS}},
  journal = {IEEE J. Sel. Top. Appl. Earth Obs. Remote. Sens.},
  year = {2025},
  doi = {10.1109/jstars.2025.3573610},
}

Loading related content...

Figure