Assessment of satellite based vegetation land surface phenology algorithms with application to a 20 year NOAA AVHRR record over Canada and Northern USA

Abstract

ABSTRACT Land Surface phenology (LSP) is related to vegetation dynamics and is an indicator of tracking surface climate change. One of the challenges in the study of LSP is the validation of satellite based LSP products . Here, we have proposed a novel methodology of the validating LSP products by applying observed temporal gap and measurement noise to representative daily NDVI reference time series from satellite imagery . Three well -known LSP algorithms (iterative Savitzky -Golay filtering - SGF [1], Asymmetric Gaussian Fitting ± AGF [2] and Logistic fitting [3, 4]) are applied to 20 years of NOAA AVHRR mea surements over biomes in Canada and Northern USA . For a given AVHRR cloud threshold, both AGF and SGF are more sensitive to the amount of gaps than to the noise in the data. Index Terms ² time seri es, phenology, comparison, validation 1. INTRODUCTION Land Surface phenology (LSP) involves the study of seasonal pattern of variation in vegetated land surfaces from remote sensing obser vations [5]. LSP is related to vegetation phenology ± the study of the timings of recurrent plant cycle events that are driven by environmental factors [6], and to changes in surface climate. Due to its significance, many attempts have been made, over the years , to produce satellite based LSP products. The Committee of Earth Observation systems (CEOS) Working Group on Calibrat ion and Validation Land Surface Phenology indicators documents a large number ( >15) of LSP products derived through the application of different algorithms to satellite measurement time series [7]. Validation of LSP algorithms is challenging d ue to differences in : i) user requirements, ii) in-situ survey methods and quantities, iii) phenology metrics iv) spatial and temporal sampling of reference measurem ents etc. Validation studies can be broadly classified into those that compare satellite based LSP with in -situ measurements [8-11 ] and those that compare satellite LSP to reference estimates derived from the same or similar satellite datasets [12 -16 ]. We focus here on the latter. Hir d and McDermid [12 ] evaluated six LSP methods over selected biomes in West -Central Alberta, Canada for their sensitivity to the amount noise. Th ey created biome specific reference sa tellite phenology time series (REF) by averaging cloud free daily MODIS NDVI time series over ~100 pixels within a biome qualitatively judged to share a similar temporal profile . Phenology estimates , derived by applying thresholds to noisy NDVI time series over a member pixel based on a selected algorithm , were compared to the same processing applied to REF . Their stu dy introduced the concept of creating a reference noise free time series from the same observations used for LSP estimat ion . Howeve r, the ir approach has limitations : i) subjective criterion for selecting inputs to reference profiles limits its usage in a global perspective , ii) noisy profiles used for validation may be biased compared to the reference , and iii) limited control on the joint pattern of gaps and noise. Kandasamy, et al. [16 ] used an ensemble of smoothers , rather than the daily average used in [12], to create a REF within a sampled biome. This approach has advantages in comparison to [12] : i) REF is less likely to be biased to a single smoother , ii) REF can be constructed even if gaps are present in the data used to create it , and iii) actual gap profiles can be applied when generating noisy measurements . They evaluated 8 candidate methods by selecting gaps from other cloud screened time series in the same biome. H owever, the no ise magnitude of the remai ning measurements was held constant. This study proposes a nove l methodology , based on [12 , 16 ] for r ealistic noise and gap simulation for validation of satellite LSP . The approach is applied to a 20 year NOAA AVHRR NDVI times series over Canada and northern USA to quantitatively evalu ate 3 well -known LSP algorithms namely, the Asymmetric Gaussian Fitting - AGF [2], Logistic Fitting - Logistic [3, 4] and iterative Savitzky -Golay filtering - SGF [1]. 2. METHODOLOGY 3522978-1-4799-5775-0/14/$31.00©2014 Crown IGARSS 2014