Estimation of soil texture by fusion of near-infrared spectroscopy and image data based on convolutional neural network

Abstract

Soil texture is very important information for various agricultural, environmental, and geological research areas, however, it has been difficult to obtain data through proximal sensing due to the lack of appropriate predictive models. In this study, sand, silt, and clay contents of soil samples were evaluated using fusion data with the near-infrared (NIR) and image based on the convolutional neural network (CNN) model. For the development of the applicable predictive model, the preprocessing methods such as standard normal variate (SNV) and derivative, image types and size were optimized, and the performance of the CNN model was enhanced using only a small amount of data (n = 250). As results, the SNV and derivative methods increased model performance (R2 > 0.91) for the NIR dataset, and the red–green–blue (RGB) and hue-saturation-value (HSV) image types showed high performance (R2 > 0.84). Meanwhile, the fusion dataset improved the model performance than the individual dataset. The model performance (R2) could be achieved over 0.93 using fusion data of size 30 × 30 regardless of soil fraction. Furthermore, this trained model showed better predictive performance (R2) than the previous models even on the test dataset: 0.69 and 0.80 for silt and clay, respectively. Therefore, these results demonstrated that the proposed CNN model using NIR and image fusion data showed the potential to accurately predict silt and clay fractions despite the small amount of training data.