Phenological models for the leaf out date in temperate and boreal Biomes determined from NDVI

1. Phenological models for the leaf out date in temperate and boreal Biomes determined from NDVI Jörg Kaduk Department of Geography, University of Leicester Sietse Los Department of Geography, University of Wales, Swansea QUEST, CLASSIC workshop on modelling plant phenology and large scale observations, 28./29.11.2005

2. Motivation Understanding and predicting the response of the terrestrial biosphere to climate variability and change requires prognostic, climate-driven canopy phenology models

3. Single species observations • Temperate and boreal trees require a period of chilling for rapid leaf budburst in spring • For single species, the heat sum required for budburst drops with increasing chilling exponentially to a threshold From Murray et al. (1989) Thermal time from 1 January to budburst (day oC>5oC) (GDD) Number of chill days (<5oC) since 1 November (NCD)

4. Questions • Do the relationships between chilling and heat accumulation carry over from single species to whole biomes? • Can consistent models be determined from large scale data? • Is there an element of adaptation to local conditions discernible? Re-evaluation and extensions of work of Kaduk & Heimann (1996), White et al. (1997) and Moulin et al. (1997) with new data and some new lessons learnt.

5. Procedure • Selection of biomes • Estimation of leaf-out date from NDVI • Calculation of GDD and NCD • Determination of models for leaf-out date using GDD and NCD

6. Data • ISLSCP II FASIR NDVI (Los et al., 2005) • GSWP-2 global 1° near surface temperature 1983-1996 (Zhao et al., 2003) • National Snow and Ice Data Center (NSIDC) weekly 25km snow cover (Armstrong et al., 2002)

7. Biomes evaluated Classification of the University of Maryland land cover map into SiB2 biomes (from ISLSCP II) 1 Evergreen Needleleaf Forest 3 Deciduous Needleleaf Forest 4 Deciduous Broadleaf Forest 5 Mixed Forests 6 Woodlands 9 Open Shrublands

8. Estimation of leaf-out date from NDVI NDVI data: Los et al. (2000), updated (10 daily, 0.5o) Four different methods tested for northern hemisphere: a. White et al. (1997) b. Largest curvature (maximal 2nd derivative of a spline fit) c. Kaduk & Heimann (1996), largest NDVI change after 30 day running mean temperature reached 5oC (Temperature data from GSWP2, daily, 1o) d. Moulin et al. (1997), using the largest increase in the change of the NDVI above a reflectance threshold.

9. a, c x Leaf-out date • Dates from methods a & c quite similar • Leaf-out dates of curvature based methods b & d quite early Mean NDVI threshold of White et al. (1997) • Snow data (Armstrong & Brodzik, 2002) suggest that curvature based methods detect snow melt instead of leaf-out d x b x • Reject methods b & d, though there might be the possibility to adjust the threshold of Moulin et al. (1997)

10. Mean leaf out date

11. Models for leaf out GDD > F* • Spring Warming Model: F*is constant, no internal state of dormancy • Sequential Model: budburst is not possible until dormancy is broken. Chilling requirement, C*, must be met before GDD accumulation • Parallel Model: GDD accumulate before C* is met, effectiveness reduced until C* is reached • Alternating Model uses a variable GDD threshold F*, F* drops with increasing chilling

12. GDD-NCD relationships Test different functional forms relating GDD, NCD and Ta 1) 2) 3) 4) 5) Versions of the alternating model Dependence of coefficients Fit with gradient method, non-linear least squares

13. Leave one out cross validation • Fit models on nine years • Test models on remaining 10th year • Avoids overfitting (fit on one year only: 10-year mean error at least 50% higher than the error for the year used for calibration) • Obtain mean absolute error estimates for model fitted on all years • Evaluate stability of model coefficients

14. Calibrated models

15. Conclusions • Difficult to apply some previously used methods to detect leaf out in temperate and boreal areas • Date of leaf out advanced ~7 days in 1980 and 1990’s • Meaningful large scale phenology models can be derived from NDVI • Some measure to avoid overfitting is required to obtain models with reasonable predictive power • Ecophysiological adaptations detectable using NDVI for large scale vegetation units

16. GDD-NCD relationships NCD and GDD at budburst for biome 9 • NCD and GDD at budburst (method 1) for biome 9 (red crosses) in 1984 and fit from function b) (black pluses) using mean winter temperature • explained variance 79% x determined from NDVI + fit from function b)