报告详情
Neuromorphic Small Object Detection for Climate Disaster Monitoring: A Bio-inspired Approach with Spatiotemporal Adaptive Learning
编号：590
访问权限：仅限参会人
更新：2025-03-31 17:50:26 浏览：2次
张贴报告
报告开始：暂无开始时间（Asia/Shanghai）
报告时间：暂无持续时间
所在会场：[暂无会议] [暂无会议段]
暂无文件
摘要
Climate disaster monitoring faces critical challenges in detecting small-scale targets (e.g., incipient wildfires, micro-scale storm cells) under complex atmospheric conditions characterized by low signal-to-noise ratios and dynamic spatiotemporal variations. To address this issue, we propose a neuromorphic small object detection framework (NSOD-CDM) that integrates bio-inspired neural mechanisms with climate data characteristics. By mimicking the human visual cortex's hierarchical processing and attention mechanisms, the NSOD-CDM architecture combines a spiking neural network (SNN) with a multi-scale feature fusion module, enabling adaptive feature extraction from multi-source remote sensing data (infrared, radar, and hyperspectral).
The key innovation lies in the event-driven dynamic learning mechanism, which utilizes spatiotemporal spike coding to suppress background noise in climate images while enhancing sensitivity to small targets (pixel area < 0.01% of image). Evaluated on the NWP-RS (Numerical Weather Prediction Remote Sensing) dataset containing 12,500 annotated disaster scenarios, our method achieves a detection precision of 92.7% for targets smaller than 32×32 pixels, surpassing conventional CNNs (78.4%) and Transformer-based models (85.9%). Notably, the neuromorphic implementation reduces computational energy consumption by 63% compared to GPU-accelerated deep learning frameworks, demonstrating feasibility for edge deployment in meteorological stations.
Case studies on typhoon eyewall localization and wildfire ignition point identification further validate the framework's robustness to cloud occlusion and illumination changes. This work provides a novel paradigm for energy-efficient and reliable climate disaster early warning systems, bridging the gap between brain-inspired computing and geoscientific applications.
The key innovation lies in the event-driven dynamic learning mechanism, which utilizes spatiotemporal spike coding to suppress background noise in climate images while enhancing sensitivity to small targets (pixel area < 0.01% of image). Evaluated on the NWP-RS (Numerical Weather Prediction Remote Sensing) dataset containing 12,500 annotated disaster scenarios, our method achieves a detection precision of 92.7% for targets smaller than 32×32 pixels, surpassing conventional CNNs (78.4%) and Transformer-based models (85.9%). Notably, the neuromorphic implementation reduces computational energy consumption by 63% compared to GPU-accelerated deep learning frameworks, demonstrating feasibility for edge deployment in meteorological stations.
Case studies on typhoon eyewall localization and wildfire ignition point identification further validate the framework's robustness to cloud occlusion and illumination changes. This work provides a novel paradigm for energy-efficient and reliable climate disaster early warning systems, bridging the gap between brain-inspired computing and geoscientific applications.
关键词
Neuromorphic Computing; Small Object Detection; Climate Disaster Monitoring; Spiking Neural Network; Edge AI;
报告人
张斌
高级工程师/教授 新疆政法学院/香港技术研究院稿件作者
全部评论
重要日期
-
会议日期
04月17日
2025
至04月20日
2025
-
04月03日 2025
初稿截稿日期
-
04月20日 2025
注册截止日期
主办单位
中国科学院大气物理研究所
承办单位
中国科学院大气物理研究所
联系方式
- 皇甫静亮
- hf******@mail.iap.ac.cn
- 133********
发表评论