Deep learning ship detection in satellite optical imagery suffers from false positive occurrences with clouds, landmasses, and man-made objects that interfere with correct classification of ships, typically limiting class accuracy scores to 88\%. This work explores the tensions between customization strategies, class accuracy rates, training times, and costs in cloud based solutions. We demonstrate how a custom U-Net can achieve 92\% class accuracy over a validation dataset and 68\% over a target dataset with 90\% confidence. We also compare a single node architecture with a parameter server variant whose workers act as a boosting mechanism. The parameter server variant outperforms class accuracy on the target dataset reaching 73\% class accuracy compared to the best single node approach. A comparative investigation on the systematic performance of the single node and parameter server variant architectures is discussed with support from empirical findings.
翻译:卫星光学图像中的深学习船舶探测存在云层、地层和人为物体的假正数,干扰了船舶的正确分类,通常将舱级精度分数限制在88 ⁇ 。这项工作探索了定制战略、舱级精度率、培训时间和云基解决方案成本之间的紧张关系。我们展示了自定义的U-Net如何在验证数据集上达到92 ⁇ 级精度,在目标数据集上达到68 ⁇ 级精度,信任度为90 ⁇ 。我们还比较了单一节点结构与参数服务器变量的比较,其工人起到促进机制的作用。参数服务器变量比目标数据集的班级精度高,达到73 ⁇ 级精度,而最佳的单节点节点方法则比较了。在经验结论的支持下,对单一节点和参数服务器变量结构的系统性能进行了比较调查。