AutoSR4EO: an AutoML Approach to Super-Resolution for Earth Observation
Abstract
Super-resolution (SR) is an image processing approach to increase the resolution of images. It is commonly used as a pre-processing step in the pipeline of a diverse range of downstream machine-learning tasks based on Earth Observation (EO) images. The manual design of SR models specific to every possible EO use case can be a laborious process, creating a bottleneck for EO analysis pipelines. In this paper, we apply Automated Machine Learning (AutoML) to automate the creation of dataset-specific SR models and present the following contributions: (i) We propose AutoSR4EO, an AutoML method for automatically constructing a neural network for SR. We design a search space based on state-of-the-art methods in SR and incorporate transfer learning from available EO datasets. Our search space can be extended with new model blocks and pre-trained weights, making it possible to adapt AutoSR4EO to future developments in the field. (ii) We evaluate the performance of AutoSR4EO based on PSNR and SSIM on 4 different datasets against that of two state-of-the-art baselines and an AutoML SR method. We evaluate two different search spaces. AutoSR4EO achieves the highest average ranking, performing well consistently over all datasets. AutoSR4EO also dominates a vanilla AutoML method, demonstrating the merit of a customised search space for the automated SR for EO images. (iii) We introduce a new real-world single-image SR (SISR) dataset, SENT-NICFI, which adds to the small collection of available multi-sensor SISR datasets.
People
- Julia Wąsala
- Mitra Baratchi
- Suzanne Marselis
- Laurens Arp
- Holger Hoos
- Nicolas Longépé
- Gurvan Lecuyer
Software
The code for AutoSR4EO and SENT-NICFI is published on GitHub, or you can download this zipfile (February 2023): [ ]
The repository also includes links to the data.
Papers
Abstract
Super-resolution (SR) is an image processing approach to increase the resolution of images. It is commonly used as a pre-processing step in the pipeline of a diverse range of downstream machine-learning tasks based on Earth Observation (EO) images. The manual design of SR models specific to every possible EO use case can be a laborious process, creating a bottleneck for EO analysis pipelines. In this paper, we apply Automated Machine Learning (AutoML) to automate the creation of dataset-specific SR models and present the following contributions: (i) We propose AutoSR4EO, an AutoML method for automatically constructing a neural network for SR. We design a search space based on state-of-the-art methods in SR and incorporate transfer learning from available EO datasets. Our search space can be extended with new model blocks and pre-trained weights, making it possible to adapt AutoSR4EO to future developments in the field. (ii) We evaluate the performance of AutoSR4EO based on PSNR and SSIM on 4 different datasets against that of two state-of-the-art baselines and an AutoML SR method. We evaluate two different search spaces. AutoSR4EO achieves the highest average ranking, performing well consistently over all datasets. AutoSR4EO also dominates a vanilla AutoML method, demonstrating the merit of a customised search space for the automated SR for EO images. (iii) We introduce a new real-world single-image SR (SISR) dataset, SENT-NICFI, which adds to the small collection of available multi-sensor SISR datasets. People
-
- Julia Wąsala
- Mitra Baratchi
- Suzanne Marselis
- Laurens Arp
- Holger Hoos
- Nicolas Longépé
- Gurvan Lecuyer
Software
The code for AutoSR4EO and SENT-NICFI is published on GitHub, or you can download this zipfile (February 2023): [ ] The repository also includes links to the data. Papers
-
- Julia Wąsala (supervisors: Dr. Mitra Baratchi & Dr. Suzanne M. Marselis & Laurens Arp &Prof.dr. Holger Hoos & Nicolas Longépé & Gurvan Lecuyer).
AutoSR-RS: an AutoML Approach to Super-Resolution for Remote Sensing Master's Thesis in Computer science at Leiden Institute of Advanced Computer Science, Leiden University, 2022.