DEEPFOREST - Exploration de l’héritage archéologique et écologique des civilisations d’Asie du Sud-Est sur les forêts tropicales par apprentissage automatique profond multimodal

Although there is ample evidence to suggest that past human activities have significantly shaped the tropical forests of South-East Asia right up to the present day, the extent and nature of this influence remain largely unknown. By assessing how ancient civilisations have impacted these ecosystems, the DEEPFOREST project aims to gain a better understanding of their current dynamics and the long-term consequences of human actions on forest landscapes. Identifying traces of past activities on a large scale and correlating them with the current state of vegetation poses a major challenge, due to the density of tropical forests and the difficulty of locating archaeological traces dating back millennia. To overcome these obstacles, the project combines a vast corpus of newly accessible remote-sensing data with complementary in situ measurements and multimodal observations — LiDAR, orthophotos, satellite imagery and hyperspectral data. These data sets are utilised to develop and train multimodal deep learning models designed to detect archaeological traces and predict key indicators and functional characteristics of forests, such as biomass, forest types and vegetation composition. The project’s interdisciplinary team, comprising experts in archaeology (EFEO), machine learning (ENPC/LIGM), forestry (AMAP-IRD) and environmental sciences (LSCE-CNRS), aims to provide new insights into these complex ecosystems and to explore the relationship between ancient settlements and current forest structures, in order to inform sustainable ecological practices.

Deliverables:

• Perron Y., Sydorov V., Wijker A.P., Evans D., Pottier C., Landrieu L., « Archaeoscape: Bringing Aerial Laser Scanning Archaeology to the Deep Learning Era », NeurIPS Benchmark & Dataset, 2024 (spotlight, top 3%) — preparatory work for the project.
• Perron Y., Sydorov V., Pottier C., Landrieu L., « Adapting Vision Transformers to Ultra-High Resolution Semantic Segmentation with Relay Tokens », Transactions on Machine Learning Research Journal, 2026. [hal-05462692]
• Perron, Y., G. Astruc, N. Gonthier, C. Mallet & L. Landrieu)., «UniverSat: Resolution- and modality-agnostic transformers for Earth observation ».
• Doctorat de Yohann Perron (ENPC), en cours depuis 2023. (submitted for publication)
• Upcoming projects: further publications planned for the period 2026–2029 (automated archaeological detection, multimodal modelling of forest indicators).