Stage M2: Spatial transfer of deep learning models for rapeseed crop mapping
6 Novembre 2023
Catégorie : Stagiaire
A 6-month master's internship at the TETIS lab (Montpellier) under the supervision of Cassio F. Dantas and Dino Ienco on domain adaptation techniques for satellite image time series data.
Detailed description at this link: https://nextcloud.inrae.fr/s/kk2bNkMHKDCQSgC
Keywords: Machine Learning, Remote sensing, Land cover mapping, Agriculture
Context and problematic:
The ever-increasing availability of remote sensing data offers the possibility to follow the evolution of a geographical area over time. The time series thus generated represent an essential source of information to efficiently manage agriculture on a territorial scale. To this end, remote sensing data is used as input to machine learning (ML) methods to provide updated land cover maps. To do so, ML methods require a large amount of ground-truth data, which poses challenges for their applicability where little or no reference data is available.
Re-using ground-truth data acquired at a particular study site to transfer the learnt model to a different area would avoid (or reduce) new costs and take advantage of previous investments. Unfortunately, directly transferring a model from a geographical zone to another one can be inefficient as the two regions may present different environmental and/or climatic conditions. This results in differences in the distribution of the acquired satellite data.
Objectives of the internship:
Developing an innovative deep learning/transfer learning method with the aim to transfer a model learnt on a particular area (where ground truth data is available) to a different geographical area where no available ground truth data is accessible.
In the context of this internship we will exploit freely available multi-temporal Sentinel-1 imagery, less sensitive to cloud occlusions due to the intrinsic nature of the SAR signal, with the aim to build a deep learning classification model for the mapping of the rapeseed crop culture  based on recent time series classification approaches [2,3]. In addition, the designed deep learning method will exploit recent domain adaptation techniques [4,5] to cope with the transfer task between three different geographical areas, namely: France, USA and Canada. We will evaluate the capability of the underlying deep learning model to be calibrated over one of the three areas and deployed on the remaining ones.
Pursuing a master's degree in computer science or signal processing with knowledge of machine learning, image analysis and data science. Knowledge in remote sensing and agriculture will be a plus.
 Maleki, S., Baghdadi, N., Dantas, C. F., et al. (2023). Artificial Intelligence Algorithms for Rapeseed Fields Mapping using Sentinel-1 Time Series: Temporal Transfer Scenario and Ground Sampling Constraints. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing (JSTARS)
 Garnot, V. S. F., Landrieu, L. (2020). Lightweight temporal self-attention for classifying satellite images time series. ECML-PKDD Workshop, AALTD.
 Foumani, N. M., Tan, C. W., Webb, G. I., Salehi, M. (2023). Improving Position Encoding of Transformers for Multivariate Time Series Classification.Preprint arXiv:2305.16642.
 Ganin, Y., Ustinova, E., Ajakan, H., et al. (2016). Domain-adversarial training of neural networks. Journal of machine learning research (JMLR).
 Zhao, H., Zhang, S., Wu, G., Moura, J. M., Costeira, J. P., Gordon, G. J. (2018). Adversarial multiple source domain adaptation. Advances in neural information processing systems.
 Nyborg, J., Pelletier, C., Lefèvre, S., Assent, I. (2022). TimeMatch: Unsupervised cross-region adaptation by temporal shift estimation. ISPRS Journal of Photogrammetry and Remote Sensing.