Adonis: Neural-enhanced Fine-grained Leaf Wetness Sensing with Efficient mmWave Imaging

Yimeng Liu, Maolin Gan, Gen Li, Yonsuk Dong, Zhichao Cao

Abstract: Predicting Leaf Wetness Duration (LWD) is crucial for plant disease control. However, the lack of standardized techniques to measure LWD precisely hampers accurate prediction. While previous works have explored various methods, they fail to quantify the actual water on the leaf, undermining their practical effectiveness and accuracy. This paper presents Adonis, an innovative approach using millimeter-wave (mmWave) radar to address the complexities of leaf wetness detection. It introduces a new metric, Leaf Wetness Level (LWL), for measuring leaf surface water. We employ advanced signal processing on mmWave signals to extract more wetness-related features in dynamic environments. Furthermore, we develop a Contrastive Learning Feature Extraction model to precisely capture wetness features and design a calibration process for the inference stage to detect LWLs accurately in real-world fields. Using a frequency-modulated continuous-wave (FMCW) radar within the 77 to 81 GHz band, Adonis is meticulously evaluated across various plants. Adonis can detect LWLs with the mean absolute error (MAE) of 4.43 in controlled environments and 6.49 in real farm conditions. The performance significantly surpasses traditional Leaf Wetness Sensors, which have an MAE of 11.84 indoors and 14.32 in field conditions. These findings have substantial implications for enhancing disease prediction and crop management.