Five-band multispectral image synthesis from UAV RGB imagery using a spectral-informed GAN for agroforestry monitoring

Vegetation indices (VI), including the Normalized Difference Vegetation Index (NDVI) and Normalized Difference Red Edge Index (NDRE), are established proxies for chlorophyll content and plant stress in tropical agroforestry systems. Computing these indices requires Red Edge and Near-Infrared (NIR) bands that standard Red-Green-Blue (RGB) cameras cannot capture. Smallholder cocoa farmers in Côte d’Ivoire, who manage around 40% of global cocoa production, operate RGB-equipped Unmanned Aerial Vehicles (UAVs) without multispectral sensors, which restricts reliable vegetation health monitoring. This study proposes SMH-GAN (Serial dual attention, Multi-head decoder, Hybrid spectral-informed Generative Adversarial Network), a modification of Pix2Next, to synthesize five-band multispectral imagery from RGB UAV data to enable VI computation without dedicated sensors. Four modifications address Pix2Next limitations. Serial Dual Attention (SDA) resolves absent band-specific feature separation by applying spectral then spatial attention sequentially. A Spectral-Informed Bottleneck (SIB) enforces vegetation reflectance constraints at the latent level without target band inputs. A Split-Head Decoder reduces gradient conflicts from joint optimization of heterogeneous bands. A Spectral Consistency Loss penalizes inter-band ratio violations to preserve VI reliability during training. SMH-GAN was evaluated on a public UAV dataset of ten cocoa agroforestry plots in Divo, Côte d’Ivoire, against Pix2Pix, TaijiGNN, and Pix2Next. The proposed method achieved mean Structural Similarity Index Measure (SSIM) of 0.876, a 4.8% improvement over Pix2Next. NDVI achieved r=0.91 with ground truth, and six indices exceeded r=0.87. These results indicate that the synthesized bands preserved the spectral relationships required for vegetation index computation on the evaluated dataset.