Lightweight Hierarchical Multimodal AI for Satellite-Edge Earth Observation: A Wildfire Detection Use Case
Abstract
Satellite remote sensing provides large-area coverage for wildfire monitoring, but timely and reliable detection remains challenging because multispectral observations can be affected by smoke, haze, cloud contamination, vegetation variability, background clutter, and sensor-dependent spectral responses. At the same time, emerging low-Earth-orbit (LEO) satellite systems in next-generation nonterrestrial networks (NTNs) increasingly require on-board intelligence to reduce unnecessary downlink load and support low-latency Earth observation (EO) services. In this article, we propose LWHM-AI, a lightweight hierarchical multimodal AI inference framework for resource-constrained satellite-edge EO, demonstrated through a wildfire detection use case. The proposed framework combines an always-executed lightweight image-only screening model with a selectively invoked multimodal verification model through an adaptive confidence gating (ACG) policy. The first stage resolves low-uncertainty samples using down-sampled multispectral inputs, while the second stage performs compact multimodal verification for ambiguous samples by fusing full-resolution image features with auxiliary meteorological and temporal context. To select the ACG operating point, we employ a Pareto-TOPSIS-based multiobjective strategy that balances detection quality and inference cost without retraining the underlying models. The framework is evaluated on three publicly available satellite wildfire datasets, namely Landsat-8, TS-SatFire, and Sen2Fire, and compared with representative lightweight and remote-sensing-oriented baseline methods. Experimental results show that the proposed architecture achieves an average F1-score above 0.92 and an average conditional inference latency of 3.6 ms under controlled workstation profiling, demonstrating a favorable relative detection-quality versus inference-cost tradeoff for time-sensitive satellite-edge EO monitoring.