Triple Cascaded Spectral Attention U-ResNet for UAV RGB-to-multispectral reconstruction in cocoa agroforestry

High-resolution multispectral imaging systems for vegetation monitoring remain costly for many smallholder farmers. Most existing deep learning approaches attempt to infer multispectral information from RGB inputs. However, such methods ignore the biophysically determined spectral order across bands of light. Conventional single-decoder designs process all spectral outputs jointly and treat homogeneous feature representations across bands as a given. This study introduces TC-SA-UResNet (Triple Cascaded Spectral Attention U-Net with ResNet Encoder) as an integrated architecture that answers the gaps left by existing methods. TC-SA-UResNet performs cascaded reconstruction from RGB to Red Edge and Near Infrared bands. The network integrates a pretrained ResNet-50 encoder, Learnable Feature Injection, Spectral Attention via Cross-Attention, and a Biophysical Spectral Consistency Loss. The framework formalizes spectral reconstruction as an optimization problem with biophysical constraints. Evaluation on UAV-acquired cocoa agroforestry imagery yields SSIM values of 0.9540, 0.8805, and 0.9233 for the Blue, Green, and Red bands, respectively. The Red Edge and NIR bands achieve SSIM scores of 0.8024 and 0.7180. The proposed method achieves the highest NIR SSIM compared to T-GAN and U-Net baselines. TC-SA-UResNet also improves Red Edge R 2 to 0.8099 and exceeds T-GAN (0.6360). The improved R 2 indicates enhanced radiometric fidelity for vegetation index computation. The reconstructed NDVI achieves a Mean Absolute Error of 0.0342 with an R 2 of 0.8160 and a Pearson correlation of 0.9068 on the test set. The reported results position the proposed framework as a credible low-cost route to multispectral monitoring for precision agriculture applications.