Bayesian Optimized ANFIS Network Using Grid Partition and Feature Spectrum for Urban Light Pollution Assessment

Abstract

Precise luminance evaluation of light pollution requires expensive and time-consuming measurement devices such as hyperspectral imaging cameras and imaging luminance-meters, which are also inconvenient to carry. To alleviate this challenge, we selected a low-cost smartphone camera sensor as an alternative tool and investigated the various factors that could influence the accuracy of the measurements by controlling the function of the sensor and assessing its performance, including light illumination conditions and device parameters. Building on this, we have developed an Adaptive Neuro-fuzzy Inference System (ANFIS) structure utilizing global Bayesian optimization and grid partitioning (GP), which integrates the advantages of fuzzy logic in handling data uncertainty with the self-learning capabilities of artificial neural networks. This method can capture the relationships between different parameter combinations while avoiding overfitting, effectively handling unseen data with rapid convergence. The experimental results demonstrate that this method reduces the error by at least 30% compared to conventional methods when tested on a dataset with previously unseen shot parameters. Using a smartphone as a measurement device offers superior portability and broader prospects compared to cameras. Leveraging the powerful processing capabilities of smartphone platforms, we can implement more advanced visualization and computational functions in the future.