Paper: DeepSpace: Super Resolution Powered Efficient and Reliable Satellite Image Data Acquisition
Authors: Chuanhao Sun, Yu Zhang, Bill Tao, Deepak Vasisht, Mahesh K. Marina
Presenter: Shiyu Liu, University of Science and Technology of China
Guest of Honor: Chuanhao Sun, Hong Kong Institute of Science & Innovation
Q: Besides being lightweight, does BSH offer other advantages over learned embeddings, local descriptors, or other modern compact feature representations?
A: The key constraint is on-board computing resources. Satellites typically only carry a Jetson GPU running in low-power mode, which makes heavy deep learning inference completely infeasible. If you train an encoder, you also need to keep it up to date to properly compress the incoming data, which adds another burden. Variational encoders are among the baselines we evaluated, and their fidelity and compression ratio are actually not as good as BSH in our setting. So while deep learning methods can certainly extract more complex features, they are simply not feasible on-board. BSH may not be optimal, but it is sufficient — and improving it would only yield marginal gains since it is just one step in the overall system.
Q: How do you ensure the reliability of the system?
A: From a systems perspective, reliability means the service remains functional even when a device fails — you can always reconstruct images with high fidelity. It does not mean the satellite or camera never goes down. Those are separate concerns.
Q: Are the expert categories and hierarchical routing structure in the MoE designed empirically, or are they supported by theoretical reasoning or data-driven evidence?
A: They are mostly empirically determined. We never claim our method is optimal. In fact, for some special regions, we can achieve compression ratios of over 1000x. The design could be further specialized — for instance, if you are targeting ocean data specifically, you could add more ocean-relevant features. The method achieves solid average performance but is not globally optimal.
Q: Which part of this research was the most challenging and time-consuming?
A: Many things in this system lack theoretical answers, which forced us into heavy empirical exploration. For example, how to quantify the redundancy of an image, how frequently to update the hash codes, how to design the hierarchical routing structure in the MoE, how many layers and how many experts to use — all of these were determined experimentally. We also only had eight GPUs at the time, which constrained the scale of experiments. These open theoretical questions — such as the most reliable method to measure image similarity in compact form, or the best way to measure image redundancy — remain unsolved. Answering them could be the foundation of an outstanding future paper.
Q: Did you conduct ablation studies on the hash code update timescale?
A: Yes. If you never update the hash code, the system performs poorly. However, between monthly and yearly updates, the difference is not that significant. The uplink bandwidth is generally not a bottleneck, so monthly or even weekly updates are feasible. From our time-series observations, there are seasonal fluctuations in the similarity level of images, which is why we settled on seasonal or monthly updates empirically. To make this choice more principled, you would need more geographical knowledge about the specific imaging region.
Q: Can Deep Space be adapted to other types of remote sensing data, such as satellite video or wireless signals?
A: It depends on whether the data exhibits repetition. For wireless signals specifically, there are scenarios — for example, estimating channel state using a combination of cloud coverage, satellite position, and radio measurements — where the patterns repeat frequently. In those cases, Deep Space could work. But for random or non-repetitive signals, it will not. Compression is only feasible where there is redundancy. For video data captured by satellite cameras, since the scenes repeat over the same locations, it could also be applicable.
Q: Did you implement a real end-to-end prototype between an actual satellite and the ground station?
A: No, that is not possible for us. We do not own a satellite. What we have is the hardware infrastructure — CPUs and Jetson GPUs — that mirrors what would actually run on board a satellite. The experiments are emulation-based. Our claim is that, at the assumed resolution and frame rate, the required downlink bandwidth is well within the capacity of LEO satellite downlinks.
Q: If the satellite had much more powerful computing resources — say, a data center on board — would the Deep Space design change significantly?
A: Yes, it would change a lot. With sufficient on-board compute, you could store images directly on board and run deep learning models for comparison and compression. One of our co-authors from UIUC has worked on exactly this kind of space-based computing scenario. But this is far from our current setting, where the constraint is a low-power edge device.
Q: If LEO satellites become dense enough to form a resource pool, could images be distributed across satellites for load balancing? And what is the inter-satellite communication bandwidth like compared to satellite-to-ground bandwidth?
A: That scenario is definitely coming if high-frequency remote sensing becomes a real business. In theory, deep enough satellite density combined with Deep Space could enable near-CCTV-like continuous coverage — tracking cars or other objects from space. However, inter-satellite communication is technically challenging. High implementation density is required, and routing and data access between satellites are open problems: you cannot guarantee immediate access to data stored on a different satellite. These are active research directions.
Q: For multi-spectral remote sensing images, does Deep Space still apply? And is there any issue specific to quantitative remote sensing tasks?
A: We did run evaluations on multi-spectral data. However, most publicly available datasets — such as those from Planet and Dynamic Earth — are primarily RGB, so we lacked sufficient multi-spectral data for a comprehensive evaluation. Empirically, we observed no significant performance difference between RGB and multi-spectral images in Deep Space, because the same location at any frequency band tends to observe similar scene patterns.
Q: Given that satellite-to-station links may face congestion, is it possible to design domain-specific congestion control for this scenario?
A: Remote sensing probably cannot afford the infrastructure cost of rerouting traffic through other satellites and base stations — that requires a fully implemented inter-satellite network, which is expensive. The congestion scenario you describe — many satellites simultaneously trying to offload to one base station — is real, but I think a more commercially viable scenario for domain-specific congestion control would be video streaming or other consumer applications, not remote sensing. The market for high-frequency commercial remote sensing does not yet really exist.
Q: What advice do you have for new graduate students who want to enter this research area?
A: A few things. First, read up-to-date papers from SIGCOMM and NSDI to understand what your peers are working on. Second, find the topic you are genuinely most interested in — for me, it was signal processing, which is why I approached this from an image compression angle rather than a pure networking angle. That different perspective was actually valuable. Third, get your hands dirty: the code and the Dynamic Earth dataset are open source, so you can start training Deep Space yourself. Finally, stay close to industry — some of my own work was done during internships at Microsoft and Samsung, and that industry proximity helped shape practical research directions.