Abstract:Monitoring soil volumetric moisture content is crucial for enhancing agricultural production efficiency and devising reasonable soil management strategies. Ultra-wide band radar, due to its high resolution and strong penetration capabilities, is widely used in real-time monitoring of dynamic agricultural soil information. However, previous processing of ultra-wide band radar signals mainly focused on time-domain features, neglecting the equally informative frequency-domain characteristics. This oversight limited the utilization of echo signals in the inversion process of soil volumetric moisture content, thereby constraining the inversion accuracy. The soil echo signals obtained from ultra-wide band radar and extracts features related to soil volumetric moisture content were preprocessed. The signals were analyzed by using shorttime Fourier transform (STFT) to investigate the time-frequency spectral characteristics related to soil volumetric moisture content variations over time. Furthermore, a soil volumetric moisture content classification and regression prediction algorithm model was established by combining these features with a convolutional neural network (CNN). Experimental results showed that based on data augmented with Gaussian white noise, the overall accuracy and Kappa coefficient for soil volumetric moisture content classification using time-frequency features combined with the CNN model were respectively 98.69% and 0.9849. Compared with support vector machine (SVM) model built with ten time-domain features and the normalized difference vegetation index (NDVI), there was an increase in overall accuracy by 21.78 percentage points and an improvement in the Kappa coefficient by 0.2515. For soil volumetric moisture content regression prediction, combining time-frequency features with a convolutional neural network regression (CNNR) model, the coefficient of determination (R2) was 0.9872, the root mean square error (RMSE) was 0.0048cm3/cm3, and the relative percent difference (RPD) was 6.2738. Compared with the CNNR model established with ten time-domain features and NDVI, there was an increase in R2 by 0.2316, a reduction in RMSE by 1.3377cm3/cm3, and an improvement in RPD by 4.2714. Overall, the method proposed showed a clear advantage over traditional signal detection and processing methods in terms of classifying and predicting soil volumetric moisture content.
