Semiterrestrial crabs evolved adaptations that allow them to stay out of the water for extended periods of time. The implication is that they experience and contribute to a rich soundscape. The project reported here aims to detect and classify sounds and vibrations produced by land crabs to accomplish two main objectives: to give insight into the acoustic environment of crabs and to build a reliable research tool that can be used in crab behavior laboratory investigations. Hence, investigators and pet owners can infer what a crab does in its laboratory or home habitat. All data was collected at the LeDuc Lab crab laboratory. The model was trained using a 1D convolutional neural network architecture. A validation accuracy of 87.3% was achieved after training on Edge Impulse. After int8 quantization, the model features only 64.1 KB flash usage and takes 237 ms for a single inference. After deployment to an Arduino Nano 33 BLE, the performance was evaluated in the same lab environment where data was collected. The results show that this project presents a useful research tool for studying semi-terrestrial crab activity.

Semisterrestrial crabs produce sound and vibrations by striking surfaces with their claws, rubbing their legs together, or using their gastric mills. Their activity in a lab environment also results in detected sounds. These activities include but are not limited to bubbling, burrowing, eating, crashing into objects, and climbing walls. This project aims to answer the following question: Can we infer crab activity from the acoustic signals they produce? The main objective was to build an artificial intelligence-enabled research tool that detects and classifies sounds and vibrations produced by crabs housed in our artificial habitats to accelerate research on semiterrestrial crabs housed in our laboratories. The flow diagram is shown in Fig. 1.

We are interested in certain crab activities, namely eating and climbing. Thus, it is reasonable only to detect and classify crab activities of interest to avoid an overwhelming continuous data stream from all the crabs housed in the lab. In the following sections, we present our approach, methods, and deployment results.

For more information, please contact the author:

lalmaghr@andrew.cmu.edu