Matthieu VERLYNDE

The aim of this thesis is to study the problem of classifying image time series in remote sensing. This is a field that raises a number of issues linked to the sheer volume of data, most of which is unlabelled. In this context, energy efficiency and frugality in data processing are crucial for two aspects (Strubell, Ganesh and McCallum 2019, Wu et al. 2022):

• The accessibility of the proposed methods to users, to enable rapid inference on an ever-increasing amount of data. This is also important for use by non-experts, who do not have extensive computational resources at their disposal.
• Reducing the ecological footprint of AI methods, which is a major challenge for sustainable development. Indeed, AI models are often very energy-intensive. The need to use classification models on data of varying spatial and temporal scales makes it crucial to develop methods that are both efficient and energy-efficient.

Recent advances in the field rely on deep neural network architectures based on convolutions that take the temporal dimension into account (Pelletier, Webb and Petitjean 2019), or with recurrent architectures coupled with attention modules (Tang et al. 2022, Sainte Fare Garnot et al. 2020, Russwurm and Körner 2018). These models are effective when a large amount of data is available, but are often ill-suited to contexts where data is scarce, or where computational resources are limited. The focus here is on developing energy-efficient classification methods.
In addition, one of the major challenges in remote sensing lies in the weakly supervised nature of remote sensing, where despite the abundance of data, reliable annotations are rare, requiring innovative techniques to make the most of the information available (Wang et al. 2020, Verbockhaven, Charpiat and Chevallier 2024), it is necessary to develop methods that are both efficient, resource-efficient and take into account the weakly supervised nature of the data. On the one hand, we need to develop model evaluation metrics that go beyond classification performance and include a trade-off with regard to energy use. Secondly, we need to develop network architectures that can reproduce high-performance classification results for the time series classification problem, while being frugal in terms of computational resources and the number of data sets used.

__________________________________________________________________________________________________________

Management: Ammar Mian, Yajing Yan (LISTIC)

• Wang et al (2020), Weakly Supervised Deep Learning for Segmentation of Remote Sensing Imagery.
• Strubell et al (2019), Energy and Policy Considerations for Deep Learning in NLP.
• Wu et al (2022), Sus- tainable AI: Environmental Implications, Challenges and Opportunities
• Pelletier et al (2019), Temporal Convolutional Neural Network for the Classification of
• Satellite Image Time Series
• Sainte Fare Garnot et al (2020), Sa- tellite Image Time Series Classification With Pixel-Set Encoders and Temporal Self-Attention.
• Russwurm et al (2018), Multi-Temporal Land Cover Classification with Sequential Recurrent Encoders
• Wang et al (2020), Weakly Supervised Deep Learning for Segmentation of Remote Sensing Imagery.
• Verbockhaven et al. (2024), Growing tiny networks: spotting expressivity bottlenecks and fixing them optimally