Dataset for Predicting Hyperspectral Signatures from Multispectral Data using cGANs for Agricultural Applications

Predicting Hyperspectral Signatures from Multispectral Data using cGANs for Agricultural Applications Authors: Aron Debruyn, Manuel Pérez-Ruiz, Stef Lhermitte, Orly Enrique Apolo Apolo This dataset is for the research on reconstructing high-resolution hyperspectral signatures from multispectral or RGB data using conditional Generative Adversarial Networks (cGANs) for agricultural applications. It provides the data needed to train, validate, and test deep learning models that converts low-dimensional spectral inputs into detailed hyperspectral curves in the 476–914 nm range. Data Collection: The dataset is derived from hyperspectral imagery captured over plots of soft wheat and barley in Seville, Spain, using a Headwall Nano-Hyperspec pushbroom camera. The data includes radiometrically corrected and preprocessed hyperspectral image patches, alongside 22 simulated multispectral/RGB input band scenarios. These scenarios are based on common agricultural sensors (like Micasense) and feature selection methods (MI, PCA), enabling the study of optimal band selection for accurate spectral reconstruction. Each T_data Numpy file has the shape of (250, 272, 60, 60) which corresponds to (amount of samples, amount of bands, height, width). Key features: High-resolution Hyperspectral Data: Contains 272 spectral (397.786–1003.180 nm) the first 35 and last 40 bands should be removed due to excessive noise to give 197 spectral bands (475.973–913.82 nm), ideal for detecting plant stress and biochemical properties. Simulated Multispectral/RGB Inputs: Includes 22 predefined input band scenarios, from simple RGB to 5-band "Centers" (Blue, Green, Red, Red-Edge, NIR) and two-band combinations, facilitating research on minimal sensor requirements. Preprocessed Image Patches: Consists of 60x60 pixel samples extracted from the original hyperspectral scans, ready for model input. Ground Truth for Validation: The original hyperspectral data serves as the target for supervised learning, allowing for precise evaluation of reconstruction accuracy using metrics like RMSE, MRAE, PSNR, SSIM, and SAM. This dataset aims to support researchers in remote sensing, computer vision, and precision agriculture by providing a benchmark for developing and validating spectral reconstruction models. By enabling the generation of hyperspectral-quality data from affordable multispectral or RGB sensors, it helps lower the barrier to advanced crop monitoring. The accompanying code for model training and evaluation can be found at: https://github.com/Aron-Debruyn/Predicting-Hyperspectral-Signatures-from-Multispectral-Data-using-cGANs.