Combination of Convolutional Neural Networks and Recurrent Neural Networks for predicting soil properties using Vis-NIR spectroscopy

摘要

Visible and Near-infrared diffuse reflectance spectroscopy (Vis-NIR) serves as a rapid and non-destructive technique to estimate various soil properties. Recently, there is a growing need for developing a more accurate and robust calibration model in large soil spectral libraries to support the implementation of effective soil quality assessments and digital soil maps at national, continental and even global scales. Traditional calibration methods, such as partial least squares regression (PLSR), support vector machines regression(SVMR), multivariate adaptive regression splines(MARS), random forests(RF), and artificial neural networks (ANN), may not be successfully applied in large spectral libraries due to their relatively weak generation performance in large regions. To overcome these weaknesses, we proposed a jointed Convolutional Neural Network (CNN) and Recurrent Neural Network (RNN) architecture called CCNVR, which combines the ability of CNN to extract the local and abstract features from the raw spectrum with the advantage of RNN to learn various dependencies of sequence features. We then compared the prediction accuracy of CCNVR with other conventional methods, namely, PLSR, SVMR, CNN, ANN, and RNN, on the selected soil properties of mineral soil samples in the Land Use/Land Cover Area Frame Survey (LUCAS) database. Of all calibration models, our proposed CCNVR achieved the best model performance with the lowest RMSE value (6.40, 0.45, 3.30, and 0.35 for OC, N, CEC, and pH, respectively) and the highest R-2 (0.73, 0.70, 0.73, and 0.86 for OC, N, CEC, and pH, respectively) for the selected properties, indicating the outstanding prediction ability of our proposed model. Besides, to quantify the robustness of different calibration models, we added different levels of white noise on the original Vis-NIR spectra of the calibration set to observe how the prediction accuracy changes in the test set. The result showed that our proposed CCNVR model has a better resistance towards noise compared to other calibration models. Finally, we explored the transferability of our proposed CCNVR model. We extended the calibration model trained on the mineral samples to the organic samples through transfer learning. The result revealed that the transfer-based CCNVR fine-tuning model had a better prediction accuracy than that of the non-transfer CCNVR model with an improvement of R-2 value from 0.79 to 0.84. The result demonstrated the excellent transferability of our proposed CCNVR model across different soil types and sample sizes.