Articles
Is it possible to obtain NDVI from RGB images?
Article number
1427_29
Pages
215 – 222
Language
English
Abstract
The normalized difference vegetation index (NDVI) is the most widely used metric for assessing the density and health of vegetation in terms of water stress, nutritional deficiency, and plant disease occurrence for precision agriculture purposes.
Usually, it is extracted from images acquired by proximal and/or remote sensing using hyperspectral and multispectral optical sensors which can often represent an unaffordable expense for smallholders.
NDVI is calculated as the normalized difference between reflectance values at visible red and near-infrared bands, thereby ranges from -1 to 1. In this study, a model based on a shallow-regressive neural network (SNN) to predict NDVI from sRGB images was developed. The model was trained on images of different plant species at different growing vegetation stages acquired by a snapshot hyperspectral camera (Specim IQ), captured in natural environmental light conditions, and then calibrated.
The model performs a pixel-to-pixel regression and shows the calibrated RGB values have a strong correlation with the NDVI, for the validation data set (r=0.91). The same shallow-regressive neural network was tested on a set of data acquired with a different non-co-registered sensor.
Specifically, the application of the model to around 1000 drone images acquired by a 6X Sentera sensor still has good performances (r=0.71). To demonstrate the reliability of the proposed method, k-means clustering (k=2) was applied to NDVI images to make a comparison between observed and predicted results evaluating r2, SSIM, and accuracy mean values. Therefore, this work shows the efficiency of the AI approach to estimate vegetation index: for the first time crop conditions could be monitored by computing NDVI using data from a low-cost RGB device. The application of this method can be an important resource for small-sized farms and a first step to making this type of technology accessible to improve production, reduce costs, and minimize environmental impact.
Usually, it is extracted from images acquired by proximal and/or remote sensing using hyperspectral and multispectral optical sensors which can often represent an unaffordable expense for smallholders.
NDVI is calculated as the normalized difference between reflectance values at visible red and near-infrared bands, thereby ranges from -1 to 1. In this study, a model based on a shallow-regressive neural network (SNN) to predict NDVI from sRGB images was developed. The model was trained on images of different plant species at different growing vegetation stages acquired by a snapshot hyperspectral camera (Specim IQ), captured in natural environmental light conditions, and then calibrated.
The model performs a pixel-to-pixel regression and shows the calibrated RGB values have a strong correlation with the NDVI, for the validation data set (r=0.91). The same shallow-regressive neural network was tested on a set of data acquired with a different non-co-registered sensor.
Specifically, the application of the model to around 1000 drone images acquired by a 6X Sentera sensor still has good performances (r=0.71). To demonstrate the reliability of the proposed method, k-means clustering (k=2) was applied to NDVI images to make a comparison between observed and predicted results evaluating r2, SSIM, and accuracy mean values. Therefore, this work shows the efficiency of the AI approach to estimate vegetation index: for the first time crop conditions could be monitored by computing NDVI using data from a low-cost RGB device. The application of this method can be an important resource for small-sized farms and a first step to making this type of technology accessible to improve production, reduce costs, and minimize environmental impact.
Publication
Authors
L. Moscovini, L. Ortenzi, S. Figorilli, S. Violino, C. Pane, V. Capparella, S. Vasta, C. Costa, F. Pallottino
Keywords
AI model, pixelwise regression, shallow-regressive neural network, vegetation index, RGB sensor, drones
Online Articles (30)