Near-Infrared Band Synthesis From Earth Observation Imagery With Learned Location Embeddings and Task-Driven Loss Functions

The acquisition of near-infrared (NIR) imagery is crucial for various remote sensing (RS) applications but is often unavailable in RS datasets, providing only the red–green–blue (RGB) bands. Most classical RS workflows rely on NIR information, such as the normalized difference vegetation index and normalized difference water index, for vegetation and soil monitoring or land cover classification workflows. To overcome this limitation, we introduce NIR-generative adversarial network (GAN), a conditional GAN designed to synthesize the NIR band directly from RGB imagery through an image-to-image translation framework. Crucially, the proposed model integrates geographical and climatic context using location embeddings, and incorporates application-specific index-derived loss functions to optimize synthesized outputs for downstream tasks. NIR-GAN enables the generation of partly synthetic datasets, facilitating the training and evaluation of multispectral machine learning models. The developed model demonstrates the capability to generate realistic NIR images from various sources of RS data, regardless of the spatial resolution and spectral characteristics of the sensor, reaching up to 26.6 dB PSNR on Sentinel-2 images. Although synthetic data will never be able to replace the acquisition of real NIR information, this approach can fill the gap in situations where this information would otherwise be absent, including but not limited to the assembly of large-scale training datasets.