Optimized ConvLSTM for Spatiotemporal Analysis of Event-Based Data in Space Surveillance Systems

Authors:

Abstract
Event-based sensors represent a novel paradigm for capturing asynchronous and high-temporal-resolution data, addressing the growing demand for efficient and precise monitoring in dynamic and resource-constrained domains. In space surveillance, where rapid object detection, accurate trajectory prediction, and real-time decision-making are paramount, these sensors provide an unparalleled advantage. In this paper, we propose an optimized ConvLSTM architecture, a novel integration of Convolutional Neural Networks (CNNs) and Long Short-Term Memory (LSTM) networks, tailored specifically to analyze the sparse and asynchronous nature of event-based data streams. By combining the spatial feature extraction capabilities of CNNs with the temporal sequence modeling strengths of LSTMs, our model captures the spatiotemporal dependencies essential for high-performance space surveillance. The architecture is meticulously fine-tuned with innovations such as dynamic pooling, an optimized learning rate schedule, and a robust event-based data preprocessing pipeline, ensuring robust performance across critical tasks. Experimental evaluations highlight the model’s exceptional performance, achieving a classification accuracy of 99.7%, along with Mean Squared Error (MSE) values of 0.0034 for velocity estimation and 0.0027 for trajectory prediction, surpassing traditional methods. These results underscore the potential of ConvLSTM in advancing space surveillance systems, offering robust capabilities for dynamic space environments.

This publication has 14 references indexed in Scilit: