Strategies and methods for predicting soil organic matter at the field scale based on the provincial near infrared spectral database

The development and provision of soil spectral library (SSL) could facilitate the application of near infrared (NIR) spectroscopy for economical, accurate, and efficient determination of soil organic matter (SOM). In this work, the performances of partial least squares regression (PLSR) and convolutional neural network (CNN) combined with the datasets of Zhejiang provincial SSL (ZSSL) and the feature subset (FS) were compared for the prediction of SOM at the target field. The FS dataset was chosen from ZSSL based on similarity to the spectral characteristics of the target samples. The results showed that compared with modelling using ZSSL, modelling using FS can greatly improve the prediction accuracy of the PLSR model, but the impact on the performance of the CNN model was limited. The method of mean squared Euclidean distance (MSD) was an effective way for determining the optimal spiking sample size for the PLSR model only using the spectral data of the spiking subset and the prediction set. The PLSR model combined with the FS dataset and the spiking subset determined by MSD achieved the optimal prediction results among all developed models, which is an accurate and easy-to-implement solution for the SOM determination based on ZSSL.