How a lightweight transformer can map cities from the sky?

Imagine flying over a city and, with an ultra‑detailed photo, spotting every building. That’s what DeepSwinLite does, a new AI model built on the Swin Transformer. Unlike massive networks that need hundreds of millions of parameters, DeepSwinLite is “light”: it uses fewer calculations and less memory while staying highly accurate. How does it work? The model combines three tricks: 1) Multi‑Scale Feature Aggregation (MSFA) gathers information at different sizes, so it recognizes both tiny houses and tall skyscrapers; 2) Multi‑Level Feature Pyramid (MLFP) blends fine details with broader concepts, improving context understanding; 3) the AuxHead provides extra supervision during training, making learning more stable. Tested on two famous datasets – Massachusetts and WHU – DeepSwinLite outperformed state‑of‑the‑art methods, reaching a global accuracy (OA) of 92.5% and an IoU of 77.9% on the former, and an impressive 98.3% OA and 92.0% IoU on the latter. After fixing ground‑truth errors in the Massachusetts data, performance rose to 94.6% OA and 79.9% IoU. Beyond the scores, the real advantage is speed: the model needs fewer FLOPs and parameters, making it perfect for smartphones or drones with limited resources. In practice, DeepSwinLite can help urban planners, rescue teams, and researchers quickly map city structures, even in emergencies, without a supercomputer.

DeepSwinLite leverages the Swin Transformer architecture, which partitions the image into local windows (window‑based self‑attention) and then combines them hierarchically. The introduced local‑global attention module enables information exchange between windows of different scales, mitigating the context loss typical of purely local transformers. The Multi‑Scale Feature Aggregation stage uses convolutions at various resolutions to build a pyramid of feature maps, ensuring patterns of varying sizes are captured. The Multi‑Level Feature Pyramid, on the other hand, implements top‑down and bottom‑up fusion, inspired by Feature Pyramid Networks, to merge high‑level semantics with spatial details. The AuxHead, placed after the encoder block, generates an auxiliary loss that regularizes gradients, speeding convergence and reducing over‑fitting. Computationally, DeepSwinLite cuts FLOPs by roughly 30% compared to Swin‑Base, thanks to fewer attention heads and a smaller embedding dimension (96 vs 128). Total parameters drop below 12 M, making the model suitable for inference on mobile GPUs or embedded CPUs. Training employed advanced data augmentation (rotations, flips, color jitter) and was optimized with AdamW (lr = 1e‑4, weight decay = 0.01). OA (Overall Accuracy) and IoU (Intersection over Union) were computed on binary building masks, showing superior robustness even on scenes with strong shadows and dense vegetation. The code is released under an MIT license, encouraging reproducibility and community extensions.