Geng Wei1,2, Mulan Zou1,2, Shuyue Wang3, Bo Liu1,2*
1School of Surveying and Geoinformation Engineering, East China University of Technology, Nanchang, 330013, China
2Jiangxi Key Laboratory of Watershed Ecological Process and Information (Platform No. 2023SSY01051), East China University of Technology, Nanchang, 330013, China
3Nanjing Real Estate Registration, Nanjing, 210001, China
*Corresponding author’s email: liubo@ecut.edu.cn
Received: 05 November 2025 / Revised: 01 March 2026 / Accepted: 09 March 2026 / Published Online: 14 March 2026
Abstract
Generating Near-Infrared (NIR) imagery from RGB spectrum offers a low-cost alternative to dedicated multispectral sensors. To investigate whether NIR imagery can be generated from standard visible-light RGB camera, visible-light RGB and multispectral NIR images over a rice field were captured by an unmanned aerial vehicle (UAV). Color space features (including HSV and CIELAB) and texture features (TF) were incorporated, and optimal model inputs were identified through feature screening method. Subsequently, four distinct models were developed for NIR image generation. Results showed that performance of generating NIR images improves substantially when HSV, CIELAB, and TF were included, and further gains were obtained after input feature selection. Pix2Pix achieved the best performance on the test dataset, with the highest determination coefficient (R²) of 0.78 and the lowest normalized Root Mean Square Error (nRMSE) of 6.41%, and the generated NIR images reached the highest Structural Similarity (SSIM) of 0.85 and Peak Signal-to-Noise Ratio (PSNR) of 28.48 dB. Moreover, feature importance analysis highlighted V, a, b, red-band contrast, green-band mean and variance as key predictors for NIR image generation. This study demonstrates a practical, low-cost approach to produce NIR imagery from standard visible-light RGB cameras, potentially reducing reliance on dedicated multispectral sensors.
Keywords: NIR generation, UAV, RGB spectra, Color space model, Texture feature
