See-Through Soil: Underground Root Tuber Sensing With RF Sensor Networks

This paper proposes a data-driven root tuber sensing (RTS) framework that uses the received signal strength (RSS) data from a radio frequency (RF) sensor network to reconstruct cross-section images of root tubers in soils. We perform extensive experiments with our data acquisition system in various environments to build a wireless potato sensing (WPS) dataset. We propose to integrate multi-branch convolutional neural networks with a diffusion neural network to enable fine-grained image reconstruction of root tubers. To deal with the multi-path effects on radio channels, we propose two domain adaptation methods: one-shot fine-tuning to update the neural network model online, and disentangled representation learning (DRL) to transfer a pre-trained model to unseen environments. Experimental results from over 1.7 million RF network measurements show the efficacy of the proposed methods across different environments. Our code and data are publicly available at https://github.com/Data-driven-RTI/MC_Diffusion.