A method combining machine learning algorithms with a new multi-sensor system to determine litter decomposition degree in forest ecosystems

Litter decomposition degree (LDD) is essential for characterizing litter decomposition process and regulating climate, material cycles and energy flows in forest ecosystems. However, very few methods or techniques are currently available to conveniently and accurately determine the LDD. In this paper, we propose a new method to determine the LDD by combining machine learning algorithms with a new self-developed multi-sensor system for simultaneous measurements of electrical conductivity (EC), volumetric moisture content (VMC), and grayscale. The new system comprises three circuit modules, which were integrated with the sensor probes in a 3D-printed measurement cartridge. Litters from four single-species stands and a mixed stand were sampled from Jiufeng National Forest Park, Beijing, China. In each stand, litters of different years were collected using litter decomposition bag method and the LDD defined as percentage of organic matter remaining (organic matter remaining/initial organic matter) was then modelled and predicted with Back propagation (BP) neural network, radial basis function (RBF) neural network, and support vector machine (SVM) algorithms based on the measurements of the EC, VMC, and grayscale. The calibration results of the new multi-sensor system showed that the relative error of the EC measurement was within 2 %, and the output of the dielectric-based frequency-domain sensor had a high correlation with the VMC of each litter with coefficient of determination R² > 0.96, and the measured and actual values of the grayscale agreed well (R² > 0.94). The predicted results of the machine learning algorithms for each sample were compared with the LDDs defined as percentage of organic matter remaining, showing that the three algorithms had the same trend as the traditional-determined LDD, with root means square error less than 0.1. In contrast, BP neural network had higher predicted accuracy in determining LDD for five kinds of forest litter samples. In general, the method proposed in this study can conveniently and accurately determine the LDD of five forest stands and has potential for wide application under field condition in near future.