Large-scale recognition of natural landmarks with deep learning based on biomimetic sonar echoes.

The ability to identify natural landmarks on a regional scale could contribute to the navigation skills of echolocating bats and also advance the quest for autonomy in natural environments with man-made systems. However, recognizing natural landmarks based on biosonar echoes has to deal with the unpredictable nature of echoes that are typically superpositions of contributions from many different reflectors with unknown properties. The results presented here show that a deep neural network (ResNet50) was able to classify ten different field sites and 20 different tracks (two at each site) distributed over an area about 40 km in diameter. Based on spectrogram representations of single echoes, classification accuracies up to 99.6% for different sites and 94.7% for different tracks have been achieved. Classification performance was found to depend on the used pulse component (constant-frequency-CF vs frequency-modulated-FM) and the trade-off between time and frequency resolution in the spectrogram representations of the echoes. For the former, classification performance increased monotonically with better time resolution. For the latter, classification performance peaked at an intermediate trade-off point between time and frequency resolution indicating that both dimensions contained relevant information. Future work will be needed to further characterize the quality of the spatial information contained in the echoes, e.g. in terms of spatial resolution and potential ambiguities.