Estimation of Soil Organic Matter in Maize Fields with Parallel Input Deep Learning Based on Vis-Nirs and Libs Fusion

Abstract

Precise estimation of Soil Organic Matter (SOM) is essential for precision agriculture. While spectroscopic methods are effective for SOM estimation, individual sensors have inherent limitations. Multi-sensor fusion provides a promising solution. This study develops a parallel input deep learning (PIDL) model that leverages visible-near-infrared spectroscopy (vis-NIRS) and laser-induced breakdown spectroscopy (LIBS) for SOM estimation. A total of 440 soil samples were collected from a maize field in Wuzhong City, Ningxia, China. Following the preprocessing of the spectra, the optimal subset of features was extracted using competitive adaptive reweighted sampling (CARS).Bidirectional long short-term memory (Bi-LSTM) and convolutional neural network (CNN) architectures were employed to construct the PIDL based on single-sensor model results. The estimation performance of different fusion strategies was compared, and the results showed that feature level fusion gave the best outcome (R2V=0.91, RMSEV=0.79, MAEV=0.74, RPDV=2.91 in validation set). This study demonstrates that multi-sensor fusion significantly improves SOM estimation accuracy. The combination of vis-NIRS and LIBS with PIDL modeling presents a highly accurate and effective method for SOM estimation, poised to become a pivotal tool in assessing soil fertility and guiding precision agricultural production.

Keywords: soil analysis, data fusion, deep learning, convolutional neural network, bidirectional long short-term memory