Regression Models between Active Sensor-Measured NDVI and UAV-Acquired Multispectral Images with Positioning Uncertainty

Abstract

Nowadays, it is frequent to monitor large crop areas using high-precision active sensors for measuring NDVI or multispectral cameras mounted on UAVs. However, the NDVI calculations using multispectral images differ from the readings of an active sensor for the same object or surface. What is more, there is a difference between the NDVI values of two multispectral images taken with different lighting conditions or height over the same object. In this paper, we propose new models to estimate NDVI using images from a multispectral camera with precision and accuracy comparable to those of a portable active sensor. For this, we propose a methodology where three lambertian reflection surfaces are chosen and characterized with a hyperspectral camera. These surfaces appear in the aerial images and are used as control points of three NDVI values in the range of 0 to 1. Then, a linear model and an exponential model, derived from the original NDVI calculation expression, are proposed and evaluated, using the spectral information of the pixels inside a region equivalent to the active sensor reflection zone and the NDVI measurements of the same sensor. After the conditioning of the data, the parameters of the models are obtained by calculating the minimum squared errors. In addition, we have done a set of tests to verify the variations of the parameters versus the positioning uncertainty, different lighting conditions and different heights in the range of 16.0 to 30.5 m. The results show that the parameters of the two proposed models vary with the height, maintaining the absolute differences of NDVI close to 0.01, which is equivalent to the resolution of the active sensor; the smallest differences occur for the linear model in the interpolation interval, while the exponential model has a better behavior near the upper limit of NDVI.