Noncontact Soil Moisture Estimation Using Continuous Wave Radar and Deep Learning

Estimating soil moisture (SM) using microwave Doppler radar is gaining attention to save fresh water and enhance crop growth and yield in agriculture due to its noncontact form of measurement. Most of the work in the literature focused on utilizing frequency-modulated continuous wave (FMCW) and ultrawideband (UWB) radar for their capability of measuring SM at different depths. However, none of the work in the literature focused on utilizing simpler architecture continuous wave (CW) radar. In this article, to estimate SM types, namely, dry, moist, and wet, a 24-GHz CW radar leveraged with time-frequency mapping is proposed. We utilized the time-frequency mapped scalogram images to train deep-learning models named DarkNet53, MobileNetV2, and ResNet101. Repetitive measurements were conducted for 108 h, capturing data from dry, moist, and wet soil samples, segmented using a 15-s window, resulting in a total of 25 920 images in the controlled environment experiment. In the outdoor environment, 12 h of data containing dry and wet samples were collected and similarly segmented, producing a total of 2880 images. Experimental results demonstrated that “DarkNet53” outperformed the other networks, achieving an accuracy of 91.3%. Additionally, with the addition of water content to the soil surface, a linear increasing trend of the phase of the radar-reflected echoes with the increase of water content was observed for short-scale study. To the best of our knowledge, this is the first reported investigation on utilizing CW radar for estimating SM, which has several potential applications, including smart irrigation systems, plant health monitoring, home plantation monitoring, and precision agriculture.