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

TL;DR

This work tackles the challenge of increasing the spatial resolution of Land Surface Temperature (LST) maps without relying on a scale-invariance assumption that can misrepresent fine textures. It introduces Scale-Invariance-Free Convolutional Neural Networks (SIF-CNN-SR1 and SIF-CNN-SR2) trained in a self-supervised manner to produce high-resolution LST that remains faithful to low-resolution observations while synthesizing texture guided by high-resolution NDVI. The optimization combines a reconstruction term with a texture term driven by a texture operator G, enabling NDVI-informed texture transfer and physical consistency. A concomitant ASTER–MODIS evaluation dataset is released, and the models are benchmarked against state-of-the-art statistical methods and a scale-invariant U-Net (SC-Unet); results show that SIF-CNN-SR1, in particular, achieves superior texture-aware performance in both spatial and Fourier domains, offering a promising path for high-resolution LST monitoring in diverse environments.

Abstract

Due to the trade-off between the temporal and spatial resolution of thermal spaceborne sensors, super-resolution methods have been developed to provide fine-scale Land SurfaceTemperature (LST) maps. Most of them are trained at low resolution but applied at fine resolution, and so they require a scale-invariance hypothesis that is not always adapted. Themain contribution of this work is the introduction of a Scale-Invariance-Free approach for training Neural Network (NN) models, and the implementation of two NN models, calledScale-Invariance-Free Convolutional Neural Network for Super-Resolution (SIF-CNN-SR) for the super-resolution of MODIS LST products. The Scale-Invariance-Free approach consists ontraining the models in order to provide LST maps at high spatial resolution that recover the initial LST when they are degraded at low resolution and that contain fine-scale texturesinformed by the high resolution NDVI. The second contribution of this work is the release of a test database with ASTER LST images concomitant with MODIS ones that can be usedfor evaluation of super-resolution algorithms. We compare the two proposed models, SIF-CNN-SR1 and SIF-CNN-SR2, with four state-of-the-art methods, Bicubic, DMS, ATPRK, Tsharp,and a CNN sharing the same architecture as SIF-CNN-SR but trained under the scale-invariance hypothesis. We show that SIF-CNN-SR1 outperforms the state-of-the-art methods and the other two CNN models as evaluated with LPIPS and Fourier space metrics focusing on the analysis of textures. These results and the available ASTER-MODIS database for evaluation are promising for future studies on super-resolution of LST.

Figures (5)

• Figure 1: Schematic representation of the U-net architecture of $\Psi_{\theta}$. Orange blocks correspond to convolutional blocks with (Conv2d, Batchnorm2d, ReLU), red blocks are average pooling reducing the size of the image by a factor two along each dimension, blue blocks are bilinear interpolation increasing the size of the image by a factor two and the green block is a two dimensional convolutional layer. The number of channels of each two dimensional convolutional layer is indicated in the figure. The input is the concatenation ($||$) along the channel dimension of the high resolution NDVI $V^{(h)}_{obs}$ and the bicubic interpolation of $T^{(l)}_{obs}$. The output is the super-resolution $T^{(h)}_{sr}$ at the spatial resoluion of $V^{(h)}_{obs}$.
• Figure 2: Center Europe area (h18v04 MODIS tile) used in this study.
• Figure 3: a) Attenuation spectra of the ASTER LST (red), LST obtained with statistical super-resolution methods TsHARP and ATPRK (blue), DMS (green), SC-Unet (black), SIF-CNN-SR1 (brown) and SIF-CNN-SR2 (orange). The attenuation spectra of the MODIS NDVI is also shown in dashed red. The curves correspond to the mean spectra over the full test dataset. The shadow around the ASTER attenuation spectra correspond to $\pm$ a standard deviation around the mean value. b) Mean Error between ASTER attenuation spectrum and the attenuation spectra of TsHARP and ATPRK (blue), DMS (green), SC-Unet (black), SIF-CNN-SR1 (brown) and SIF-CNN-SR2 (orange). The curves correspond to the mean error over the full test dataset and the shadowed areas correspond to one standard deviation around the mean. The red horizontal line indicates the zero.
• Figure 4: For one random image from the validation dataset, visualization of the LST of MODIS and ASTER respectively at 1km and 250m of spatial resolution alongside the super-resolution LST obtained with the different approaches.
• Figure 5: Statistical analysis of the image visualized in figure \ref{['fig:vis0']}. a) Boxplots and Violinplots representing the statistical distribution of the values of the high pass filtered LST (see equation \ref{['eq:highpass']}) of ASTER and obtained with the different super-resolution approaches. b) Attenuation spectra of the ASTER LST (red), LST obtained with statistical super-resolution methods TsHARP and ATPRK (blue), DMS (green), SC-Unet (black), SIF-CNN-SR1 (brown) and SIF-CNN-SR2 (orange). The attenuation spectra of the MODIS NDVI is also shown in dashed red.