Very High Resolution satellite and aerial imagery are used to monitor and conduct large scale surveys of ecological systems. Convolutional Neural Networks have successfully been employed to analyze such imagery to detect large animals and salient features. As the datasets increase in volume and number of images, utilizing High Performance Computing resources becomes necessary. In this paper, we investigate three task-parallel, data-driven workflow designs to support imagery analysis pipelines with heterogeneous tasks on high performance computing platforms. We analyze the capabilities of each design when processing 3097 and 1575 images for two distinct use cases, for a total of 4,672 satellite and aerial images and 8.35 TB of data. We experimentally model the execution time of the tasks of the image processing pipelines. We perform experiments to characterize resource utilization, total time to completion and overheads of each design. Our analysis shows which design is best suited to scientific pipelines with similar characteristics.

超高分辨率卫星和航拍图像用于监测和进行生态系统的大规模调查。卷积神经网络已成功用于分析此类图像以检测大型动物和显着特征。随着数据集数量和图像数量的增加，利用高性能计算资源变得必要。在本文中，我们研究了三种任务并行、数据驱动的工作流设计，以支持高性能计算平台上具有异构任务的图像分析管道。我们分析了每种设计在处理两个不同用例的 3097 和 1575 幅图像时的能力，总共有 4,672 幅卫星和航拍图像以及 8.35 TB 的数据。我们通过实验模拟图像处理管道任务的执行时间。我们进行实验来表征每个设计的资源利用率、完成总时间和开销。我们的分析显示了哪种设计最适合具有相似特征的科学管道。