More Optimized Stochastic Configuration Networks (MOSCNs) Enabled Smart Sensing for Soil Carbon Content Quantification

Stochastic Configuration Networks (SCNs) have the ability to maintain the universal approximation property by employing a constructive scheme for entire neural networks and applying a supervisory mechanism by randomly assigning input weights and biases. A few studies investigated the optimal parameters of random distribution for assigning weights and biases precisely. However, they still lack further illustration of more optimal random distributions and generalization exploration on the real-world datasets. Therefore, in this paper, we present more optimal stochastic configuration networks (MOSCNs) by applying fractional terms to randomly assign input weights and biases for fine-grained SCN construction. Furthermore, we studied optimal parameters for up to seven common random distributions using microwave reflectance signals from different soil carbon content levels to validate the MOSCNs’ generalization. Results revealed that using fractional SCNs achieved around 4% classification accuracy improvement compared with general SCNs and around 20% classification accuracy benefits compared to general classifiers.