Numerical Modeling of the Stability of Color Features of Grapevine Leaves under Changing Natural Illumination

Abstract

This paper addresses the problem of improving the stability of color features of agricultural crop leaves under changing natural illumination conditions in RGB camera-based monitoring. The relevance of the study is determined by the dependence of recorded color-channel values on the intensity and spectral composition of solar radiation, imaging geometry, sensor characteristics, and atmospheric conditions. Such variability may reduce the reproducibility of automated plant-image analysis in field conditions. A numerical model of digital camera signal formation is proposed, taking into account the spectral distribution of solar radiation, the spectral reflectance of the leaf surface, the spectral sensitivities of sensor color channels, atmospheric parameters, and solar position. Model calculations were performed for the Krasnodar Krai during daylight hours. Two grapevine leaf conditions were considered: healthy and diseased. For each condition, color features were calculated in the RGB, HSV, and CIE Lab color spaces, and their variability under changing illumination conditions was evaluated. The mean standard deviation of color-feature components was used as a stability indicator. The preservation of diagnostic information was assessed using the standardized interclass separability coefficient and the normalized distance between class centroids. The original data and three classical color normalization methods—Gray World, White Patch, and Shades of Gray—were compared. The results show that transforming RGB values into the HSV and CIE Lab color spaces alone does not ensure invariance of color features to changes in natural illumination. Within the adopted model, the Gray World method provides the lowest temporal variability of features and the highest interclass separability between healthy and diseased leaves among the considered methods. The obtained results can be used to support the selection of image preprocessing methods in the design of automated agricultural crop monitoring systems.