From RGB to NIR: Predicting of near infrared reflectance from visible spectrum aerial images of crops

Masi Aslahi, Kevin G. Stanley, Ian Stavness, Hema Duddu, Steve Shirtliffe, Sally Vail, Kirstin Bett, Curtis Pozniak

Abstract:

Near infrared spectroscopy (NIR) provides rich information in agricultural operations and experiments to determine crop parameters which are not visible to the human eye. Collecting the NIR spectral band requires a multispectral camera which is typically more expensive and has lower resolution than a comparable RGB camera. We investigate image-to-image translation as a means to generate an NIR
spectral band from an RGB image alone in aerial crop imagery. Aerial images were captured via a multispectral sensor mounted on an unmanned aerial vehicle (UAV) flown over canola, lentil, dry bean, and wheat breeding trials. A software workflow was created to preprocess raw aerial images creating a dataset suitable for training and evaluating deep learning based band inferencing algorithms. Two different experiments including in-domain and out-of-domain experiments over different crop types in our dataset were conducted to evaluate efficacy in an agricultural context.