Abstract
<title>Abstract</title> <p>Coronal holes are recurrent source regions of high-speed solar-wind streams, and their evolution is relevant to both coronal physics and space-weather monitoring. Future high-cadence EUV imagers will produce full-disk data volumes that may exceed the telemetry available for downlink, particularly for deep-space or otherwise bandwidth-limited missions. On-board region-of-interest (ROI) selection is a practical means of reducing the volume of full-resolution data. This paper evaluates two FPGA-oriented architectures for real-time localization of dark coronal-hole regions in 4096 × 4096 SDO/AIA 19.3 nm images. The integral-image pipeline with sliding window (IIP-SW) uses block averaging, disk masking, thresholding, and integral-image scoring of candidate windows. The ultra-low-memory orthogonal scanning method (ULS-OS) estimates the approximate center of the most prominent dark region using row- and column-wise run-length logic without performing full connected-component labeling. Synthesized for a Kintex UltraScale KU060-class device at 100 MHz, both cores use no BRAM in the reported implementation. On the COSPAR-ISWAT community benchmark dataset, ULS-OS achieves the lowest IoU variance among the tested methods, a median IoU of 0.761, and an additional algorithmic latency of 2.56 µs. A June 2025 coronal-hole high-speed-stream (CH-HSS) interval is included as a case study: the ROI localized by ULS-OS is compared with near-Earth solar-wind (OMNI) and geomagnetic (Kp) observations to examine how hardware-level localization may support source-region tracking and retrospective analysis of space-weather events. Both methods return an ROI center based on computationally efficient proxy metrics (axial span or local density) rather than exact maximum-area labeling, striking a balance between hardware cost and localization stability.</p>