Neural network detection of small moving radar targets in an ocean environment

Small icebergs and pieces of icebergs are virtually undetectable with conventional marine radar systems. The authors describe a detection scheme for such icebergs. The scheme uses the chirplet transform, a wavelet-inspired transform, to generate images of the Doppler-shifted radar returns from icebergs and ocean surfaces. The images are classified using a neural network trained with the backpropagation algorithm, incorporating weight sharing and optimal brain damage paradigms. The network's architecture is motivated by the known physiology of animal vision. The network design incorporates temporal information. Performance has surpassed the benchmark Fourier-based detection scheme.<>