On-Orbit DNN Distributed Inference for Remote Sensing Images in Satellite Internet of Things

In satellite Internet of Things (IoT), the remote sensing satellites capture images and then transmit them to a ground station through low Earth orbit (LEO) communication satellites for model inference. However, this process results in significant transmission latency and communication overhead. In response, researchers have proposed various satellite on-orbit model inference methods. Nonetheless, the limited computation capacity and memory space of a single remote sensing satellite impose processing delays when dealing with large quantities of high-resolution images, thereby making it difficult to ensure real-time service. To tackle this issue, we propose an on-orbit deep neural network (DNN) distributed inference framework for remote sensing images in satellite IoT, leveraging the availability of numerous LEO computing satellites. Designing such a framework involves two crucial questions: first, determining which LEO satellites should participate in distributed DNN inference, and second, how to partition the images among the selected LEO satellites. To address these questions, we formulate the distributed inference process as a mixed integer nonlinear optimization problem, which is known to be NP-hard. The objective is to minimize overall energy consumption while ensuring that the distributed inference is accomplished when the satellite dynamic network remains unchanged. We initially propose a dynamic optimization algorithm that derives the optimal solution with rigorous theoretical guarantees. Subsequently, to reduce computational complexity, we introduce an approximate solution based on an improved simulated annealing algorithm. We demonstrate that the approximate algorithm performs within a limited range of the optimal algorithm. Finally, we build a heterogeneous testbed based on Kubernetes and conduct extensive experiments to validate that our proposed algorithms reduce energy consumption by an average of 24.63% and 25.98% on the Faster-RCNN inference model, 47.09% and 47.51% on the RetinaNet inference model, and 53.36% and 48.08% on the Yolov5 inference model on the two datasets compared to the baselines.