Geospatial models has a strange problem. It can recognise a house in one crop, then forget the road network around it a few kilometres later. That’s because most remote sensing models are still built around small image tiles. They label patches of land, one crop at a time. A building is a building. A road is a road. A field is a field. But satellite imagery doesn’t work like a folder of neatly cropped photos. It comes as huge scenes covering hundreds of square kilometres, where the meaning of a pixel often depends on what surrounds it. A narrow strip of water could be a river, canal, drainage channel, or pond edge. A pale rectangle could be a roof, greenhouse, road surface, or bare ground. The local texture gives clues, but the wider geography often gives the answer. This creates a bit of a trade-off. Use small crops and the model keeps sharp detail, but loses context. It can see the road surface, but loses the road network. It can label water pixels, but loses the shape that tells you whether it’s a pond, river, lake, or canal. Use the full image and the model gets the broader scene, but fine detail gets compressed. Narrow roads blur. Small buildings disappear. Boundaries get messy. A new paper from Beihang University and NTU tries to solve this with SFR-Net, a model for ultra-wide area remote sensing segmentation. The core idea is pretty simple: make the model look at the same place from multiple “altitudes” at once. For each target area, SFR-Net creates three aligned views. A local view for fine detail. A short-range view for nearby context. A long-range view for the wider landscape. All three are centred on the same location. The model isn’t stitching together random tiles. It’s building a stack of views around one place, closer to how a person might move between a drone image, a city map, and a regional map. The authors call this a scale-frustum representation. Then the model fuses the views in stages. First, the local view absorbs nearby context. Then that richer view absorbs the broader scene. Instead of choosing between detail and context, it builds from one into the other. The results are meaningful. On GID, SFR-Net reaches 74.67% mIoU and 86.94% overall accuracy, beating the previous best by 1.72 percentage points in mIoU. On FBPS, the harder dataset with 24 fine-grained land-cover classes, it reaches 77.24% mIoU and 92.91% overall accuracy. That’s a 4.29 point mIoU gain over the previous best. That second result is the more interesting one. Fine-grained land-cover mapping is where the confusion gets worse: river versus pond, road versus bare ground, small building versus surrounding urban fabric. The model improves most where geography starts doing the work that texture can’t. Remote sensing models have borrowed heavily from normal computer vision. That helped the field move fast, but aerial imagery has a different structure. Roads, rivers, forests, cities, and fields are spatial systems. Their meaning depends on scale, shape, continuity, and surroundings.