Reconstruction of spectral irradiance in a real application with a multi-channel spectral sensor using convolutional neural networks.

Lighting is not only a key mediator for the perception of the architectural space but also plays a crucial role regarding the long-term well-being of its human occupants. Future lighting solutions must therefore be capable of monitoring lighting parameters to allow for a dynamic compensation of temporal changes from the optimal or intended conditions. Although mostly based on synthetic data, previous studies adopting small, low-cost, multi-band color sensors for this kind of parameter estimation have reported some promising preliminary results. Building up on these findings, the present work introduces a new methodology for estimating the absolute spectral irradiances of real-world lighting scenarios from the responses of a 10-channel spectral sensor by using a convolutional neural network approach. The lighting scenarios considered here are based on a tunable white floor lamp system set up at three different indoor locations and comprise combinations of LED, fluorescent, tungsten, and daylight lighting conditions. For white light mixtures of the various spectral components, the proposed reconstruction methodology yields estimates of the spectral power distribution with an average root-mean-square error of 1.6%, an average Δu'v' of less than 0.001, and an average illuminance accuracy of 2.7%. Sensor metamerism is discussed as a limiting factor for the achievable spectral reconstruction accuracy with certain light mixtures.