Detection of autumn leaf phenology and color brightness from repeat photography: Accurate, robust, and sensitive indexes and modeling under unstable field observations. (November 2019)

• Record Type: Journal Article
• Authors: Koide, Dai; Ide, Reiko; Oguma, Hiroyuki
• Is Part Of: Ecological indicators. Volume 106(2019)
• Publication Date: 2019-11
• Subjects: AOI area of interest -- DNs digital numbers -- LDA linear discriminant analysis -- QDA quadratic discriminant analysis -- MDA multinomial discriminant analysis -- RPRT recursive partitioning and regression tree model -- SVM support vector machine -- RF random forest model -- SDs standard deviations; Color brightness -- Phenology -- Alpine vegetation -- Betula ermanii -- Sorbus matsumurana -- Color indexes -- Discriminant models; Environmental monitoring -- Periodicals; Environmental management -- Periodicals; Environmental impact analysis -- Periodicals; Environmental risk assessment -- Periodicals; Sustainable development -- Periodicals; 333.71405
• DOI: 10.1016/j.ecolind.2019.105482
• Languages: English
• ISSNs: 1470-160X

Abstract

Highlights: Methods for assessing autumn leaf phenology and color brightness were compared. Unintentional observation error was mimicked by random resampling. For color duration, quadratic or multinomial discriminant analysis with RGB is best. Spline-fitted VARI and RGR indexes are suitable for peak DOY detection. Brightness analysis should use different indexes depending on color and objective. Abstract: Autumn leaf phenology and its color brightness provide valuable information for managing forest carbon cycles and cultural ecosystem services. Digital repeat photography has provided standard phenological data, but the methodologies for detecting autumn leaf coloring have various strengths and weaknesses. We assessed the accuracy, sensitivity, and uncertainty of various model and color index combinations for detecting autumn leaf coloring. Then we identified the most robust and sensitive methods, using digital repeat photography data from Japanese alpine vegetation. For determining autumn leaf color duration, quadratic or multinomial discriminant analysis using RGB digital numbers had the highest accuracy (hit ratio > 0.7). For determining the peak day of autumn leaf color and its color brightness, we compared uncertainty of methodologies by randomly resampling 80% of the data 20 times to mimic observation errors (e.g., due to heavy rain). The spline-fitted red/green reflectance ratio (RGR) and visible atmospherically resistant index (VARI) proved robust for detecting the peak day (median SD = 1.25). Uncertainty of color brightness was also low when using VARI fitted by a double logistic model for both red and yellow leaves (median coefficient of variation = 1.03). These two indexes are stable despite atmospheric effects, which may result in robustness to daily variation in conditions (e.g., fog). We compared sensitivity of leaf color brightness: RGR and excess red (ExR) fitted by a double logistic model had the highest sensitivity to red and yellow leaves exceeding the average of other combinations by 26% and 88% in median values, respectively. The small denominator or lack of a denominator of these indexes increases the sensitivity to red or yellow. Our results demonstrate the averaged accuracy, sensitivity, and robustness of each methodology among our research sites with different camera observations. These methods should help in utilizing hidden big data from web cameras or past photos that were not intended for scientific research to properly assess autumn leaf phenology and its color brightness.