Session
Science/Mission Payloads Research & Academia
Location
Salt Palace Convention Center, Salt Lake City, UT
Abstract
We present a multi-modal detection and attribution algorithm for maritime domain awareness designed to execute on low-power FPGA-based edge compute hardware. Spaceborne hyperspectral sensors generate data volumes far exceeding practical downlink capacity, yet the onboard compute platforms available for real-time processing operate within strict power, memory, and radiation constraints. The proposed algorithm fuses RX anomaly scoring, ACE matched filtering, multi-spectral index evaluation, wake-aware spatial scoring, and spectral glint cancellation in a five-stage pipeline designed for hardware-efficient execution. Mixed-precision quantization assigns FP16 arithmetic to covariance-sensitive stages and INT8 to threshold-based spectral computations, achieving a 12.4× model size reduction and 6.1× compute reduction relative to a floating-point baseline while retaining 99.1% of FP32 anomaly detection performance. The algorithm is mapped to the LACE-C3A edge computing system of the ITASAT2 CubeSat, operating within an 8W payload budget and 256 MB DDR memory constraint, and is evaluated on hyperspectral imagery from the GHOSt-04 commercial constellation. The pipeline delivers vessel detection and maritime anomaly attribution (oil-slick classification and vessel wake identification) within a single satellite overpass window. Provisional validation against partially labeled scenes yields 66.7% vessel recall at the current operating point, with a staged validation roadmap targeting operational acceptance.
Document Type
Event
Multi-Modal Detection and Attribution Algorithm for Low Power Edge Compute Hardware
Salt Palace Convention Center, Salt Lake City, UT
We present a multi-modal detection and attribution algorithm for maritime domain awareness designed to execute on low-power FPGA-based edge compute hardware. Spaceborne hyperspectral sensors generate data volumes far exceeding practical downlink capacity, yet the onboard compute platforms available for real-time processing operate within strict power, memory, and radiation constraints. The proposed algorithm fuses RX anomaly scoring, ACE matched filtering, multi-spectral index evaluation, wake-aware spatial scoring, and spectral glint cancellation in a five-stage pipeline designed for hardware-efficient execution. Mixed-precision quantization assigns FP16 arithmetic to covariance-sensitive stages and INT8 to threshold-based spectral computations, achieving a 12.4× model size reduction and 6.1× compute reduction relative to a floating-point baseline while retaining 99.1% of FP32 anomaly detection performance. The algorithm is mapped to the LACE-C3A edge computing system of the ITASAT2 CubeSat, operating within an 8W payload budget and 256 MB DDR memory constraint, and is evaluated on hyperspectral imagery from the GHOSt-04 commercial constellation. The pipeline delivers vessel detection and maritime anomaly attribution (oil-slick classification and vessel wake identification) within a single satellite overpass window. Provisional validation against partially labeled scenes yields 66.7% vessel recall at the current operating point, with a staged validation roadmap targeting operational acceptance.
