Towards a NAS Benchmark for Classification in Earth Observation

Results of this project are explained in the final report.



Recently, vast efforts in the field of Automated Machine Learning (AutoML) and Neural Architecture Search (NAS) have been invested in the creation of performance-based databases [Ying2019], gathering large sets of models and their performances for various learning tasks. Since the resulting benchmarks act as simulators, they have enabled NAS researchers and practitioners to reduce the cost of deployment of search algorithms via simple cost-free queries, for various use-cases, e.g., natural language processing and computer vision. While the existing benchmarks are beneficial for their respective applications, how does NAS differ on remote sensing data compared to classical computer vision? After analyzing such benchmarks with Earth observation (EO) data, the latest evidence [Traoré2022] shows that “classical” computer vision DL models do not work “out-of-the-box” on remote sensing problems, thus new tailored architectures are required. To address this problem, we propose to create the first NAS database for classification in EO, a step towards EO-specialized neural network architectures.

The project has two goals

  1. Propose a methodology based on fitness landscape analysis for assessing the quality of a surrogate NAS database.
  2. Prepare such EO NAS database, for the So2Sat LCZ42 [Zhu2020] use case (Sentinel-1 and -2 data).

For this project, Python and a DL framework (e.g., TensorFlow, PyTorch) will be used to develop the solution. The code will be run on an HPC environment using SLURM. A technical report should be written using LaTeX. All tools can be used remotely.


  • [Traoré2022] Traoré, K.R., Camero, A. and Zhu, X.X., 2022. Landscape of Neural Architecture Search across sensors: how much do they differ?. In XXIV ISPRS Congress - The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences.
  • [Ying2019] Ying, C., et al., 2019, May. NAS-Bench-101: Towards Reproducible Neural Architecture Search. In International Conference on Machine Learning (pp. 7105-7114). PMLR.
  • [Zhu2020] Zhu, X.X., Hu, J., Qiu, C., Shi, Y., Kang, J., Mou, L., Bagheri, H., Haberle, M., Hua, Y., Huang, R. and Hughes, L., 2020. So2Sat LCZ42: A benchmark data set for the classification of global local climate zones [Software and Data Sets]. IEEE Geoscience and Remote Sensing Magazine, 8(3), pp.76-89.

Accepted students to this project should attend (unless they have proven knowledge) online workshops at the LRZ from TBA. More information will be provided to students accepted to this project.