Abstract
Accurate estimation of soil gravimetric moisture is essential for the efficient management of water resources in precision agriculture, particularly in semi-arid regions. This study proposes a convolutional neural network (CNN) to estimate soil gravimetric moisture using multispectral data acquired by unmanned aerial vehicles (UAVs). The research was conducted on a clay-loam soil plot in Guasave, Sinaloa, Mexico, during different post-irrigation drying phases. A DJI Mavic 3 Multispectral drone equipped with sensors in the Green (560 nm), Red (650 nm), Red-Edge (730 nm), and NIR (860 nm) bands was employed, capturing imagery with a spatial resolution of approximately 4 cm/pixel. In situ gravimetric measurements were collected to validate the performance of the developed CNN model. The proposed CNN architecture consisted of three Conv1D layers with 32 filters, one MaxPooling layer, a flattening layer, and a fully connected dense layer, specifically designed to capture complex nonlinear relationships between spectral bands and soil moisture. The model achieved outstanding performance, with a mean squared error (MSE) of 0.0058, a coefficient of determination (R²) of 0.91, and a mean absolute error (MAE) of 0.046 on the test dataset. Predictions were highly correlated with actual measurements (r = 0.92, p < 0.001). Statistical analyses confirmed homoscedasticity and residual normality, indicating model robustness. Furthermore, the CNN-based approach outperformed traditional empirical indices, demonstrating high sensitivity to within-field microvariability.

This publication has 19 references indexed in Scilit: