Soil PH Measurement Based on Compressive Sensing and Deep Image Prior

Abstract

Soil quality is vital in agriculture. People often use sensor networks to obtain the soil data of a piece of land. Sometimes, people detect soil data by using one-dimensional (1-D) ultrawideband (UWB) signals, which is too energy consuming. Hence, we want an energy-saving model for this condition. Compressed sensing (CS) is a feasible model to save energy. As soon as CS was put forward, it attracted wide attention. However, how to design a suitable sparse dictionary is still a problem until now. Aiming at different situations, designers should change their sparse dictionary. Unfortunately, soil data are always changing because of the variance of weather and environment. Therefore, if there is a computational intelligent CS algorithm which is suitable for variant signals, the problem is solved. In this paper, we proposed a deep learning model of CS, which can avoid designing a sparse dictionary. We combine deep learning and CS together and regard the output of networks as the recovery signals, because deep learning's end-to-end structure solves the above-mentioned problem. We use deep image prior, which is derived from deep convolutional generative adversarial networks to recover 1-D signals of soil data. Meanwhile, we used the mean square error between the recovery signals and the original signals of soil echoes to verify our algorithm. Finally, we classified PH values with recovery signals based on Xgboost because of its excellent classification ability. The experiment results show that our model's accuracy is much higher than those by traditional CS algorithms with a relatively small number of sensors. Also, our model has a better robustness.