Deep fusion approach: Combining hyperspectral imaging and ground penetrating radar for accurate cornfield soil moisture mapping

Precise estimation of soil moisture is critical for refining irrigation strategies, improving crop production, and ensuring efficient management of water supplies. This study employs a time-series L-pika hyperspectral sensor, covering the spectral range of 400–1100 nm, and an air-launched Zond Aero 500 GPR sensor, capturing data across various corn growth stages. The research focuses on assessing soil moisture at different soil depths, with an emphasis on accuracy of 10 cm and 30 cm of the root zone. While GPR provides reliable measurements at near-surface depths, its effectiveness decreases at lower levels. In contrast, the canopy spectrum relates closely to root zone water content, aiding in the estimation of soil moisture through a model informed by canopy spectra. Integrating these datasets aims to address the limitations of each method and enhance the accuracy of soil moisture measurements throughout the crop's lifecycle. To analyze the spectral signals from the canopy and the amplitude signals from the GPR, two separate one-dimensional convolutional neural network (1D-CNN) networks were developed. The features extracted by these networks were then utilized in Gradient Boosting Machine (GBM) and Artificial Neural Network (ANN) models to estimate soil moisture. The 1DCNN-ANN model demonstrated superior performance, particularly at soil depths of 10 cm and 30 cm. At 10 cm, it achieved an impressively low RMSE of 1.7 % and a high R^2 of 0.82, indicating precise predictive accuracy. The model also minimized bias to nearly zero, effectively balancing overestimations and underestimations across the datasets. At the deeper layer of 30 cm, the ANN outperformed the XGBoost model, showing robust performance with an RMSE of 2.9 % and an R^2 of 0.79. Moreover, the improvement in Lin’s Concordance Correlation from 0.84 to 0.90 for the integrated approach between the models suggests that the ANN more accurately reflects the variability in the true data, enhancing the reliability of the predictions. These outcomes highlight the effectiveness of the 1DCNN-ANN in handling complex datasets and its consistent success in refining model accuracy and reliability.