Automated detection and tracking of space debris using Explainable AI

人类在空间的活动和干预，例如报废的卫星、发射的火箭，导致空间积累大量碎片。据报道，太空中存在超过30，000块大于10厘米的空间碎片和超过1亿块大于1毫米的碎片。除此之外，一些自然空间环境组成部分（如轨道上的流星体）导致污染程度迅速上升，反过来又干扰了可持续的空间运行。总的来说，这种污染对目前正在使用的遥感卫星的预期/正常运行造成很大干扰。影响力：强有力的避免碎片机动在减少卫星之间、卫星之间以及卫星与闲置物体之间相互勾结的风险方面发挥着至关重要的作用。差距：数据收集以及精确探测和跟踪碎片轨迹是昂贵和困难的，即使对于地面设施也是如此;在有限的空间监视设施内实现这一目标更具挑战性，我们的首要目标是通过管理大规模图像库来区分无碎片空间环境图像和包含碎片的图像，并对碎片类型进行准确分类。在这一步骤之后，将开发用于跟踪多类碎片的先进算法，例如神经网络辅助卡尔曼网。目标：建立一个大型空间环境图像库，其中包括不同类别的空间碎片;利用深层神经网络开发碎片探测模型;对多种类别的碎片实施跟踪算法，克服遮挡和背景杂乱的挑战。(vi)探索可解释的人工智能技术，为深度学习模型添加可解释的功能，以便决策对用户透明。预测和跟踪算法必须在推理时间内快速执行，以产生准确的结果。提高跟踪速度对于实时碎片跟踪模型来说尤为重要。

Human activities and interventions in space e.g. dead satellites, launched rockets lead to accumulation of a large amount of debris in the space. It is reported that more than 30,000 of pieces of space debris that are larger than 10 cm and more than 100 million pieces of debris larger than 1 mm exist in the space. In addition to that, some natural space environment components (e.g. meteoroids in orbits) lead to a rapid rise in the level of pollution and in return perturb the sustainable space operations. Collectively, this pollution leads to a great interference to the expected/normal operation of currently active remote sensing satellites. Impact: Robust debris avoidance manoeuvres play a vital role in reducing the risk of collusion between satellites, with each other and also between satellites and idle objects. Gap: Data collection as well as detection and tracking of the trajectories of debris accurately is costly and difficult, even for ground-based facilities; which is even more challenging to achieve it within the limited space surveillance facilities.Our first aim is to distinguish the debris-free space environment images from the ones including debris and classifying the type of debris accurately by curating a large-scale image repository. Following this step, advanced algorithms e.g. NN-aided KalmanNet, for tracking multiple classes of debris will be developed. This involves overcoming the challenge of occlusion and background clutter.Objectives:(i) To curate a large image repository of space environment including different categories of space debris; (ii) to develop debris detection models using deep neural networks, (ii) to implement tracking algorithms for multiple classes of debris that overcome the challenge of occlusion and background clutter. (vi) To explore explainable AI techniques to add explainable functionality to the deep learning models so that the decision is transparent to the users. Prediction and tracking algorithms must perform quickly during the inference time to yield accurate results. Enhancing tracking speed is especially imperative for real-time debris tracking models.