Concluding Discussion

We build Serval, a distributed computation framework for near-realtime insights from Earth imagery satellites. Serval can deliver latency sensitive imagery such as forest fire imagery in minutes as opposed to hours or days of delay for traditional in-order delivery systems. We conclude by listing some possible extensions of Serval in future work:

• Multi-modal imagery: Serval currently focuses on RGB imagery captured by satellites. Increasingly, satellites capture other forms of imagery such as radar, hyperspectral, and multi-spectral. We believe Serval can naturally extend to support these emerging image types.

• Merging filters across queries: As the number of applications scales, there are more opportunities to reduce redundant compute within and across queries. Multiple queries may share filters to reduce compute on satellites. Moreover, multiple neural network models may share weights for a subset of the layers and present opportunities for model merging techniques like [48] to optimize compute on the satellite.

• Architectural optimizations: We did not consider architectural optimizations such as model pruning or precision drop to pack more compute on the limited satellite resources. Such techniques can further improve Serval’s performance.

我们构建了Serval，一个用于从地球影像卫星中获取近实时洞察的分布式计算框架。对于如森林火灾影像等延迟敏感数据，Serval能够在数分钟内完成交付，而传统的按序传输系统则需要数小时乃至数天的延迟。最后，我们列出Serval在未来工作中一些可能的扩展方向作为总结：

• 多模态影像（Multi-modal imagery）： Serval当前专注于卫星捕获的RGB（可见光）影像。然而，越来越多的卫星开始捕获其他形式的影像，如雷达、高光谱和多光谱影像。我们相信Serval可以自然地扩展以支持这些新兴的影像类型。

• 跨查询的滤波器融合（Merging filters across queries）： 随着应用数量的扩展，在查询内部及跨查询之间减少冗余计算的机会也随之增多。多个查询可以共享滤波器，以减少卫星上的计算量。此外，多个神经网络模型可能在部分网络层上共享权重，这为利用像[48]这样的模型融合技术来优化星上计算提供了机会。

• 架构优化（Architectural optimizations）： 我们尚未考虑如模型剪枝或精度降低等架构优化方法 ，以在有限的卫星资源上容纳更多计算任务。这类技术可以进一步提升Serval的性能。