Inversing soil salinity in cotton fields using spectroscopy sample augmentation and moisture correction

Rapidly inverting soil salinity is crucial to the soil water and salt migration for the prevention of the secondary salinization. However, the accuracy and efficiency of soil salinity inversion models are hindered to the coupling relationship between water and salt in soil, particularly for time-consuming and labor-intensive soil collection. This study aims to reduce the interference of soil moisture for obtaining better sample diversity and further improving the robustness of soil salinitu inversion models by using spectral technology. A total of 113 normal and 115 saline soil samples were collected in the Xinjiang cotton fields. These samples were further subjected to different levels of wetting treatment. and subsequently 467 soil samples with varing salt and moisture contents were obtained. Soil salt content was calibrated using the conductivity of soil leaching solution. Spectral data of samples was captured using an ASD ground object spectrometer (400-1000 nm) and a near-infrared spectrometer (960-1693 nm). The soil moisture was also corrected using the external parameter orthogonalization (EPO). Additionally, deep convolutional generative adversarial networks (DCGAN) with different transposed convolution stride strategies were designed to evaluate the sample set using Fréchet Inception Distance (FID) scores. Machine learning models were employed to invert the soil salinity, including partial least squares regression (PLSR), random forest (RF), and one-dimensional convolutional neural network (1D-CNN) models using VGG (VNet), EfficientNet (ENet), and ResNet (RNet) architectures. The results demonstrated that the EPO can effectively reduce the interference of moisture on the salinity, indicating the better prediction performance of different models. RNet out performed PLSR, RF, VNet, Enet, RNet, and exhibited the best performance to predict the soil salinity in cotton fields. The lightweight residual neural network without attention mechanism was more suitable for one-dimensional hyperspectral data. There was an increase in the convolution stride and kernel length of the deep convolutional adversarial generative network. The better samples were obtained for the hyperspectral data with long sequences. The superior FID scores were achieved in the generated augmented sample set using Generator B (designed with the larger convolution stride and kernel size), compared with the rest. Specifically, the FID scores were reduced by 7.9% and 13.4%, respectively, compared with GA and GC. The weight distribution of attention was optimized after expanding the training set by DCGAN. The stability and accuracy of the model were further enhanced to predict the soil salinity under certain constraints on training samples. The EPO-DCGAN(GB)-RNet (called EPO-DCGAN-RNet) model was achieved in the superior RMSE and R² values of 136.472 μS/cm and 0.910, respectively, on the validation set, compared with the EPO-SG-RNet (using SG filtering denoising) and EPO-RNet (without sample augmentation). Furthermore, 1D-CNN with Grad CAM was employed to identify the characteristic bands of soil conductivity in the soil leaching solution of cotton field. In summary, an accurate inversion model EPO-DCGAN-RNet was constructed for inversing soil salinity in the cotton fields using spectral technology. Water correction and sample augmentation were incorporated for the soil salt composition. The improved model has the promising potential to the salt-tolerant cotton varieties and irrigation strategies using slightly salty water in cotton fields.