NDVI and Beyond: Vegetation Indices as Features for Crop Recognition and Segmentation in Hyperspectral Data

Abstract

Vegetation indices have long been central to vegetation monitoring through remote sensing. The most popular one is the Normalized Difference Vegetation Index (NDVI), yet many vegetation indices (VIs) exist. In this paper, we investigate their distinctiveness and discriminative power in the context of applications for agriculture based on hyperspectral data. More precisely, this paper merges two complementary perspectives: an unsupervised analysis with PRISMA satellite imagery to explore whether these indices are truly distinct in practice and a supervised classification over UAV hyperspectral data. We assess their discriminative power, statistical correlations, and perceptual similarities. Our findings suggest that while many VIs have a certain correlation with the NDVI, meaningful differences emerge depending on landscape and application context, thus supporting their effectiveness as discriminative features usable in remote crop segmentation and recognition applications.

Keywords: NDVI; classification; hyperspectral imaging; remote sensing; similarity metrics; vegetation indices.

Figure 1

PRISMA hyperspectral image visualization on the date of 23 March 2024.

Figure 2

Examples of UAV-HSI hyperspectral images, rendered as RGB color components using a neural network-based method for visualizing multisource remote sensing data by mapping spectral inputs to RGB [43].

Figure 3

Class (pixel) distribution in the UAV-HSI dataset. X axis—crop type. Y axis—percentage of pixels with the respective crop.

Figure 4

The masks over the PRISMA image that have been used for zone segmentation.

Figure 5

Histograms of the considered vegetation indices with respect to the considered zone. These results suggest that there is potential in using vegetation indices as discriminative features.

Figure 6

The statistical and imaging similarity of the selected vegetation indices with respect to the NDVI for various categories of landscape (agricultural land, urban, forest, pasture). The figure is better viewed in digital form, under zoom-in.

Figure 7

Plot of pixels in (DVI, NDVI) space for 4 types of vegetation considered. The data is more similar for “urban”, while for “forest” the points are uncorrelated.

Figure 7

Plot of pixels in (DVI, NDVI) space for 4 types of vegetation considered. The data is more similar for “urban”, while for “forest” the points are uncorrelated.

Figure 8

Wilcoxon values when testing for similarity between a VI and NDVI. Values smaller than 0.05 indicate disimilarity. The maximum computed value is ${10}^{-218}$.

Figure 9

Accuracies for Random Forest under different hyperparameter (in-bag ratio and number of features considered in a node) settings.

Figure 10

Accuracies for the Support Vector Machine under different hyperparameter (cost and $\gamma$ parameter of the Gaussian kernel considered in a node) settings.

Figure 11

Accuracies for the Multilayer Perceptron under different hyperparameter (number of hidden layers and learning rate) settings.

Figure 12

Accuracies for the different learning rates ($\lambda$) in the context of Gradient Boosting Machine.

Figure 13

Example of pixel-wise independent prediction using Random Forest.

Figure 14

Feature importance computed as the impact of a split based on the feature in trees for Random Forest and Gradient Boosting Machine. Higher scores mean the feature is more important. While the NDVI obviously dominates, the rest are comparable.