Abstract
Satellite-borne optical sensors provide a promising means for global air-traffic monitoring, yet their observations are often fragmented and limited in temporal resolution, making reliable trajectory forecasting highly challenging. Here we present HiFormer, a direct multi-step deep learning framework specifically designed for trajectory prediction under sparse space-based observations. The framework integrates convolutional, recurrent, and attention-based sequence modeling within a unified architecture, enabling the capture of short-term maneuvers, medium-range motion trends, and long-range dependencies in a single forward process. To address the lack of suitable datasets, we construct a large-scale synthetic benchmark of 12,000 trajectories representing four canonical motion patterns, and compile 1000 fragmented ADS-B flight segments covering diverse global routes. Extensive experiments demonstrate that HiFormer reduces multi-step prediction errors by up to 30% on synthetic data and 10% on real-world ADS-B tracks compared with representative baselines. These results establish HiFormer as a robust framework for space-based air-traffic monitoring and highlight its potential for forecasting tasks across other domains with sparse and irregular observations.