UAV imaging for shallow-water habitat mapping: Comparing performance between hyperspectral, multispectral and RGB imaging

Summary

Unmanned aerial vehicles (UAVs) have become an increasingly popular platform for mapping shallow-water habitats such as kelp forests, seagrass meadows, and coral reefs. Most commercial UAV imaging systems have RGB (red-green-blue) sensors, and some have multispectral imagers (MSI), typically with 4-8 wavelength bands in the visible and infrared range. Hyperspectral imagers (HSI) have hundreds of spectral bands, and can also be operated via UAVs, but at a much higher cost. This study uses a hyperspectral dataset to simulate multispectral and RGB datasets, and compares the habitat classification accuracy across the three modalities. The original dataset is a set of 199 hyperspectral images from four coastal areas along the Norwegian coast. HSI, MSI and RGB datasets were used to train a "U-Net" neural network, using the exact same architecture for each modality. RGB and MSI attain weighted F1-scores of 0.784 and 0.798, respectively, making them quite similar in overall accuracy. HSI attains a weighted F1-score of 0.857, and clearly performs better than the two other modalities, especially for sand, maerl and brown algae .