Lighting Spectrum Optimization With Deep Learning for Moss Species Classification

Mosses, due to their sensitivity to environmental changes, are utilized in investigations related to air pollution, water quality, and carbon consumption and emissions. Methods modeling vegetation distribution, such as species distribution modeling (SDM), are useful for understanding climate change, but require precise sampling, which is a significant effort when large-scale periodic surveys are required. Remote sensing technology using satellite imagery and unmanned aerial vehicles (UAVs) reduces labor, but it is difficult to determine the species that are surrounded by a canopy in forests. Additionally, adequate lighting is crucial for detailed imaging in dark environments lacking ambient light. Hence, we propose a method for obtaining spectral information on moss in the forest using a deep learning model to train convolutional neural network models while optimizing a suitable light source for moss identification. For the image classification tasks of five species of Sphagnum, an optimized light source that combines 10 different spectral distributions within 400–800 nm, excluding spectral information at 600–700 nm from the RGB image’s red channel and emphasizing that at 700–800 nm, improved discrimination accuracy by 10% compared with that of images obtained with the D65 sunlight source. Imaging using lighting spectrum optimization is more cost-effective than multispectral imaging, and provides more accurate classification of plant species compared with RGB images. By using six or more optimal light sources, the classification accuracy was almost equivalent to that of a spectral bands selection model with spectral-wise self-attention from hyperspectral imaging (HSI) with 400–800 nm in 20 nm step (21 dimensions). The proposed method, which does not require expensive spectral cameras, has the potential to deliver recognition performance comparable to HSI, making it a practical and cost-effective solution for moss species identification.