Enabling High-Speed, High-Resolution Space-based Focal Plane Arrays with Analog In-Memory Computing

We present a novel space-based remote sensing architecture design that uses analog in-memory computing (IMC) to efficiently embed neural network processing in an 8K×8K, 40 kHz focal plane array. Convolutional and recurrent neural networks are developed to detect weak point-like events amid clutter and to be robust to analog device errors. Accurate neural network event detection, amplitude estimation, and localization are demonstrated on an analog IMC chip based on charge trap memory, and the response to space radiation is modeled.