Evaluation of Parallelism Strategies in Remote Sensing Image Processing Pipelines: A Study with the EuroSAT Dataset

The efficient processing of large-scale remote sensing digital images is a problem of optimization of computational resources with direct applications in agricultural and environmental monitoring. This work experimentally evaluates three parallelism strategies applied to a digital image processing pipeline: sequential execution, parallelism via multiprocessing and vectorization with NumPy. The experiments were conducted on the publicly available EuroSAT dataset, using a subset of 2,000 Sentinel-2 satellite images with ten land use and land cover classes. The required pipeline includes reading, grayscale, Gaussian filtering, edge detection by Canny, and saving the processed digital images. The results are analyzed in the light of Amdahl's Law, with speedup and parallel efficiency calculation for each configuration tested. The data obtained allows estimating the intrinsic sequential fraction of the pipeline and identifying the bottlenecks that limit the gain with the use of parallelism.