Applying deep learning to large-scale geospatial imagery presents a fundamental trade-off: high-resolution datasets are simply too large to analyse in a single pass. While downsampling preserves broader spatial context, it sacrifices fine-grained detail. Conversely, slicing imagery into discrete tile grids introduces arbitrary boundaries that bisect objects and isolate features from their surroundings. Consequently, downstream detection and segmentation models receive truncated context, compromising inference quality. While overlapping tiles offer a partial mitigation, they merely shift the burden to complex post-processing required to resolve conflicting predictions.

In this talk, we present a novel solution to bridge tile boundaries using the latest vision foundation models like DINOv3. DINOv3 processes raw imagery into uniform token grids, where each patch embedding captures rich, high-dimensional spatial context without relying on human annotations. By comparing these embeddings, semantically and spatially coherent patches can be clustered across an image. We demonstrate how this unsupervised clustering capability extends beyond individual tile boundaries to encompass arbitrarily large regions. This approach reduces reliance on tile overlaps and lossy downsampling, equipping downstream detection and segmentation models with the unbroken context needed to accurately delineate complex geographical features.

To highlight its practical value within geospatial workflows, we demonstrate our Pytorch pipeline and showcase its performance against a diverse range of satellite and aerial imagery datasets.