Modelo para la reconstrucción hiperespectral a partir de imágenes RGB en agricultura de precisión

Abstract

Near-infrared (NIR) imagery is essential for vegetation monitoring; however, capturing NIR typically requires specialized cameras, increasing costs in precision agriculture. This work presents a comparative analysis of five deep learning models for reconstructing the NIR band from RGB images: HSCNN-D, SR-UNet, SRAWAN, SRBFWU-Net, and a cGAN. All models were trained and evaluated on a common dataset of paired RGB-NIR agricultural images, with performance assessed using RMSE, MRAE, MAE, PSNR, and NDVI accuracy. Our results show that SRBFWU-Net achieved the best reconstruction fidelity, obtaining the lowest errors (MRAE 0.40092, RMSE 0.13157, MAE 0.10706) and highest PSNR (17.65 dB), along with NDVI consistency (NDVI MAE 0.17121; low Pearson R values across models indicate limited fine-grained correlation). HSCNN-D followed closely, demonstrating robust performance through its deep dense architecture. On the one hand, SR-UNet presented an appropriate balance between simplicity and effectiveness. On the other hand, SRAWAN and cGAN presented higher errors despite their complex designs. These results show that using deep learning to reconstruct near-infrared (NIR) data from RGB images is a viable method for producing accurate NDVI maps. This approach aims to facilitate low-cost crop monitoring using standard cameras. In addition, it provides recommendations for selecting models based on different factors such as accuracy, efficiency, and practical deployment considerations in precision agriculture.