Soil organic matter content prediction in tobacco fields based on hyperspectral remote sensing and generative adversarial network data augmentation

Soil organic matter (SOM) is a key indicator of soil health and fertility. Although hyperspectral remote sensing combined with deep learning can replace traditional laboratory analysis to achieve rapid SOM prediction, field collection of samples is limited by terrain resulting in data scarcity, which severely restricts the ability of model generalization. To this end, the study proposes a doubly regularized Wasserstein generative adversarial network (DR-WGAN-GP), which solves the overfitting problem caused by small samples by fusing a multi-objective loss function to achieve joint hyperspectral-SOM data augmentation. A quantitative metric to quantify the quality of data generated during data augmentation model training using machine learning model accuracy is also proposed. In addition, three SOM prediction models, namely, elastic network (ENet), partial least squares regression (PLSR), and one-dimensional convolutional neural network (1D-CNN) were built to compare and analyze the effects on model accuracy before and after data augmentation. The experimental results show that: the performance of the prediction model after data augmentation is obviously improved, DR-WGAN-GP improves the performance of the prediction model more significantly compared with the traditional WGAN-GP, and the best prediction effect is achieved by combining with 1D-CNN, and the R2, RPD, and RMSE of the validation set are up to 0.86, 2.67, and 2.64, respectively, which is about 13.51 % compared with that of the best model built on the original modeling set (R2, RPD, and RMSE of the validation set are 0.76, 2.03, and 3.48, respectively), the R2 improvement is about 13.51 %. This study can realize the demand of limited spectral sample set augmentation and provide a new idea for accurate prediction of SOM under a small sample set.