This HPC compute request supports the nAIture Capital Accounting project, funded by EFRE (European Fund for Regional Development, EFRE-20801764). One of the project's milestones is to develop an AI-based system for automated tree-cadastre analysis across North Rhine-Westphalia (NRW), with downstream applications in urban planning, industry, and forest management. The accuracy and generalisability of every downstream component rest on the quality of a self-supervised pre-training stage, which is the focus of this request: training high-capacity vision encoders on four-band (RGB + near-infrared) aerial imagery of NRW at a ground sampling distance of 10 cm. At this resolution, fine-grained canopy structures, individual tree crowns, and species-discriminative texture patterns become distinguishable — but only a sufficiently powerful encoder, trained at scale, can reliably extract them.
The central scientific contribution is a novel dual-stream vision encoder pre-trained with an enhanced, JEPA-style (joint-embedding predictive) algorithm. Dual-stream architectures are the established means of containing the extreme computational cost of ultra-high-resolution imagery, and the joint-embedding predictive paradigm — itself inherently dual-stream — is a natural fit for pre-training them.
We evaluate this method rigorously against a comprehensive set of established baselines (MAE with ViT, SimMIM with Swin, and I-JEPA with ViT), each pre-trained from scratch on the identical corpus of 1.3 million image tiles, so that performance differences reflect the method rather than the data or its scale. The trained weights of all models will be published openly. Because the encoder is trained at the finest available resolution, it transfers readily to downstream aerial and satellite tasks at coarser resolutions and in other regions (e.g. the FLAIR-Hub and URUR benchmarks), making it a broadly reusable asset for large-scale environmental monitoring worldwide.
Principal Investigator, Company and Country
Moritz Weiß, University of Wuppertal, Germany