Day 4

Detailed paper information

Back to list

Paper title Benchmarking deep learning models for land cover classification with Sentinel-2 imagery
Authors
  1. Ioannis Papoutsis National Observatory of Athens Speaker
  2. Nikolaos Ioannis Bountos National Observatory of Athens
  3. Angelos Zavras National Observatory of Athens
  4. Dimitrios Michail Harokopio University of Athens
Form of presentation Poster
Topics
  • C1. AI and Data Analytics
    • C1.04 AI4EO applications for Land and Water
Abstract text One prominent application for remote sensing (RS) imagery is land use / land (LULC) cover classification. Machine learning, and deep learning (DL) in particular have been widely adopted by the community to address LULC classification problems. A particular problem class is multi-label LULC scene categorization that is set-up as a RS image scene classification problem, with DL showing excellent performance for such Computer Vision tasks.

In this work we use BigEarthNet, a large labeled dataset based on single-date Sentinel-2 patches for multi-label, multi-class LULC classification and rigorously benchmark DL models analysing their overall performance under the light of both speed (training time and inference rate) and model simplicity with respect to LULC image classification accuracy. We put to the test state-of-the-art models, including Convolution Neural Networks (CNN), Multi-Layer Perceptrons, Vision Transformers, EfficientNets and Wide Residual Networks (WRN) architectures.
In addition, we design and scale a new family of light-weight architectures with very few parameters compared to typical CNNs, based on Wide Residuals Networks that follow the EfficientNet paradigm for scaling. We propose a WideResNet model enhanced with an efficient channel attention mechanism, which achieves highest f-score in our benchmark. With respect to a ResNet50 state-of-the-art model that we use as a baseline, our model manages 4.5% higher averaged f-score classification accuracy for all 19 LULC classes, and is trained two times faster.

Our findings imply that efficient lightweight deep learning models that are fast to train, when appropriately scaled for depth, width and input data resolution, can provide comparable and even higher image classification accuracies. This is especially important in remote sensing where the volume of data coming from the Sentinel family but also other satellite platforms is very large and constantly increasing.

Papoutsis, I., Bountos, N.I., Zavras, A., Michail, D. and Tryfonopoulos, C., 2021. Efficient deep learning models for land cover image classification. arXiv preprint arXiv:2111.09451.