FPGA Acceleration of Multi-Temporal Change Detection on High Resolution Images

Abstract

The Earth Observation domain is rapidly flourishing thanks to the adoption of big data technologies. This is possible thanks to efficient data storage and processing infrastructures, but most importantly thanks to the development of data analytic applications with machine learning techniques. Several applications have been developed that perform change detection on Earth Observation satellite images with the help of machine learning models. Examining changes of a designated area over a period of time produces a big amount of data leading to demanding requirements in terms of fast access, storage and computation. In this diploma thesis, we focus on such an application, that creates generic change detection maps for pairs of time-consecutive Sentinel-2 data products that represent exactly the same field of view. The goal of the thesis is to improve the response time of the tool without loss in accuracy. For this purpose, we leverage the computing power of heterogeneous resources, targeting an Intel Strarix 10 FPGAs and utilize the OpenCL High Level Synthesis framework to create an efficient accelerator. Our study first performs application profiling in order to identify the performance bottlenecks. We then focus our acceleration efforts on optimizing the bottlenecks with the help of built-in HLS tool optimization techniques as well as architectural and algorithmic optimizations. We employ both fine-grain and coarse grain parallelism and explore the design space to identify architectures optimized towards different objectives, i.e both performance and throughput. The accelerator is integrated into the original python application and evaluated over real Sentilel-2 images. The best architecture delivers an overall speedup of x7 over the software baseline.