VIE-Net: Regressive U-Net for Vegetation Index Estimation

• Published in: IEEE access Vol. 13; pp. 144650 - 144661
• Main Authors: Capparella, Valerio, Nerio Nemmi, Eugenio, Violino, Simona, Costa, Corrado, Figorilli, Simone, Moscovini, Lavinia, Pallottino, Federico, Pane, Catello, Mei, Alessandro, Ortenzi, Luciano
• Format: Journal Article
• Language: English
• Published: IEEE 2025
• Subjects: Accuracy; Agriculture; AI-powered sensors; Cameras; end-to-end approach; Indexes; machine learning application; Meters; Neural networks; Normalized difference vegetation index; open source; regressive convolutional neural network; remote and proximal sensing; RGB sensor; Sensors; Training; Vegetation mapping
• ISSN: 2169-3536; 2169-3536
• DOI: 10.1109/ACCESS.2025.3598124

Vegetation indexes (VIs) are important indicators in agriculture, revealing valuable information about the vegetative status of crops through nondestructive evaluation methods. Among these indexes, the Normalized Difference Vegetation Index (NDVI) is a key metric used for assessing plant cover and health by combining Near-Infrared (NIR) and Red reflectance. NDVI calculation is based on multispectral cameras equipped with NIR sensors. However, the presence of this sensor is what makes the device costly and therefore impractical for small-scale farms. To address this limitation, recent works have explored the use of artificial intelligence to build AI-powered RGB cameras as a more affordable alternative for NDVI estimation. This has been done by means of generative artificial intelligence (often prone to hallucinations) or via shallow neural networks (pixel-wise regression) with the drawback of a high computational cost. Here, we introduce an end-to-end non-generative approach for NDVI estimation from calibrated RGB images. The proposed model, called VIE-Net, is a convolutional neural network based on a regressive version of the U-Net architecture. The model is tested on two datasets with images captured at 25 m above ground level (remote sensing) and 1 meter from the subject (proximal sensing), achieving correlation performance up to <inline-formula> <tex-math notation="LaTeX">r^{2} = 0.98 </tex-math></inline-formula> when non-vegetative background is removed. A lightweight version of the model was also tested, achieving <inline-formula> <tex-math notation="LaTeX">r^{2} = 0.84 </tex-math></inline-formula>. This approach not only provides a cost-effective solution for NDVI estimation but also improves the reliability of vegetation health assessment using standard RGB images.