|
Diffusion-guided lightweight remote-sensing semantic models for on-board AI
|
Authors: Xinyue Li, Hongwei Li, Jingwen Wang, Chunhui Wang, Chaojie Zhang
Status: Final
Date of publication: 30 September 2026
Published in: ITU Journal on Future and Evolving Technologies, Volume 7 (2026), Issue 3, Pages 440-451
Article DOI : https://doi.org/10.52953/FHNP6360
|
Abstract:
Earth observation missions increasingly benefit from interpreting imagery close to the sensor, but high-capacity semantic models are difficult to operate under embedded compute, memory, and power constraints. This paper presents a diffusion-guided semantic model compression framework for resource-constrained on-board AI. During ground training, a frozen pre-trained diffusion U-Net supplies multi-scale structural priors. Feature, attention, and response-level knowledge is then transferred to compact student networks; only the selected student and task head are retained for deployment. The task head combines channel-spatial attention with class imbalance-aware supervision, and the distilled model is converted through ONNX and TensorRT for FP16/INT8 execution. On ISPRS Vaihingen, the ResNet-34 U-Net student obtains 71.33% mIoU and 83.21% mF1, versus 71.70% and 83.73% for the 566.09-M-parameter teacher. Its 22.1M parameters and 40.3 GFLOPs correspond to reductions of 96.10% and 98.94%, while throughput on an RTX 3090 increases from 1.2 to 32 FPS. On Jetson AGX Orin under the 15 W profile, TensorRT INT8 occupies 22.2 MB, reaches 22 FPS at 12.5 W, retains 70.74% mIoU, and improves energy efficiency from 0.61 to 1.76 FPS/W. The full diffusion representation also records the smallest mean relative mIoU drop (20.62%) under haze, shadow, and mixed sensor noise. These results support diffusion-to-student knowledge transfer as a model-compression path for on-board Earth observation semantic analysis. |
Keywords: Earth observation, embedded deployment, knowledge distillation, low-bit inference, model compression, on-board AI, semantic representation, space computing
Rights: © International Telecommunication Union, available under the CC BY-NC-ND 3.0 IGO license.
|
|
| ITEM DETAIL | ARTICLE | PRICE | |
|---|
ENGLISH 
Full article (PDF) |
|
| 0
| Free of charge | DOWNLOAD |
|
|
| |