Signal Processing for Remote Sensing - Chapter 8

8 Relating Time-Series of Meteorological and Remote Sensing Indices to Monitor Vegetation Moisture Dynamics J. Verbesselt, P. Jonsson, S. Lhermitte, I. Jonckheere, J. van Aardt, and P. Coppin CONTENTS 8.1 Introduction....................................................................................................................... 151 8.2 Data..................................................................................................................................... 153 8.2.1 Study Area............................................................................................................. 153 8.2.2 Climate Data.......................................................................................................... 154 8.2.3 Remote Sensing Data ........................................................................................... 155 8.3 Serial Correlation and Time-Series Analysis................................................................ 156 8.3.1 Recognizing Serial Correlation........................................................................... 156 8.3.2 Cross-Correlation Analysis ................................................................................. 158 8.3.3 Time-Series Analysis: Relating Time-Series and Autoregression................. 160 8.4 Methodology...................................................................................................................... 161 8.4.1 Data Smoothing..................................................................................................... 161 8.4.2 Extracting Seasonal Metrics from Time-Series and Statistical Analysis...... 163 8.5 Results and Discussion.................................................................................................... 164 8.5.1 Temporal Analysis of the Seasonal Metrics..................................................... 164 8.5.2 Regression Analysis Based on Values of Extracted Seasonal Metrics......... 165 8.5.3 Time-Series Analysis Techniques....................................................................... 166 8.6 Conclusions........................................................................................................................ 167 Acknowledgments..................................................................................................................... 168 References ................................................................................................................................... 168 8.1 Introduction The repeated occurrence of severe wildfires, which affect various fire-prone ecosystems of the world, has highlighted the need to develop effective tools for monitoring fire-related parameters. Vegetation water content (VWC), which influences the biomass burning processes, is an example of one such parameter [1–3]. The physical definitions of VWC vary from water volume per leaf or ground area (equivalent water thickness) to water mass per mass of vegetation [4]. Therefore, VWC could also be used to infer vegetation water stress and to assess drought conditions that linked with fire risk [5]. Decreases in VWC due to the seasonal decrease in available soil moisture can induce severe fires in ß 2007 by Taylor & Francis Group, LLC. most ecosystems. VWC is particularly important for determining the behavior of fires in savanna ecosystems because the herbaceous layer becomes especially flammable during the dry season when the VWC is low [6,7]. Typically, VWC in savanna ecosystems is measured using labor-intensive vegetation sampling. Several studies, however, indicated that VWC can be characterized temporally and spatially using meteorological or remote sensing data, which could contribute to the monitoring of fire risk [1,4]. The meteorological Keetch–Byram drought index (KBDI) was selected for this study. This index was developed to incorporate soil water content in the root zone of vegetation and is able to assess the seasonal trend of VWC [3,8]. The KBDI is a cumulative algorithm for the estimation of fire potential from meteorological information, including daily maximum temperature, daily total precipitation, and mean annual pre-cipitation [9,10]. The KBDI also has been used for the assessment of VWC for vegetation types with shallow rooting systems, for example, the herbaceous layer of the savanna ecosystem [8,11]. The application of drought indices, however, presents specific operational challenges. These challenges are due to the lack of meteorological data for certain areas, as well as spatial interpolation techniques that are not always suitable for use in areas with complex terrain features. Satellite data provide sound alternatives to meteorological indices in this context. Remotely sensed data have significant potential for monitoring vegetation dynamics at regional to global scale, given the synoptic coverage and repeated temporal sampling of satellite observations (e.g., SPOT VEGETATION or NOAA AVHRR) [12,13]. These data have the advantage of providing information on remote areas where ground measurements are impossible to obtain on a regular basis. Most research in the scientific community using optical sensors (e.g., SPOT VEGETA-TION) to study biomass burning has focused on two areas [4]: (1) the direct estimation of VWC and (2) the estimation of chlorophyll content or degree of drying as an alternative to the estimation of VWC. Chlorophyll-related indices are related to VWC based on the hypothesis that the chlorophyll content of leaves decreases proportionally to the VWC [4]. This assumption has been confirmed for selected species with shallow rooting systems (e.g., grasslands and understory forest vegetation) [14–16], but cannot be generalized to all ecosystems [4]. Therefore, chlorophyll-related indices, such as the normalized difference vegetation index (NDVI), only can be used in regions where the relationship among chlorophyll content, degree of curing, and water content has been established. Accordingly, a remote sensing index that is directly coupled to the VWC is used to investigate the potential of hyper-temporal satellite imagery to monitor the seasonal vegetation moisture dynamics. Several studies [4,16–18] have demonstrated that VWC can be estimated directly through the normalized difference of the near infrared reflect-ance (NIR, 0.78–0.89mm) rNIR, influenced by the internal structure and the dry matter, and the shortwave infrared reflectance (SWIR, 1.58–1.75mm) rSWIR, influenced by plant tissue water content: NDWI ¼ rNIR ÿrSWIR (8:1) NIR SWIR The NDWI or normalized difference infrared index (NDII) [19] is similar to the global vegetation moisture index (GVMI) [20]. The relationship between NDWI and KBDI time-series, both related to VWC dynamics, is explored. Although the value of time-series data for monitoring vegetation moisture dynamics has been firmly established [21], only a few studies have taken serial correlation into account when correlating time-series [6,22–25]. Serial correlation occurs when data collected through time contain values at time t, which are correlated with observations at ß 2007 by Taylor & Francis Group, LLC. time t–1. This type of correlati on in time-se ries, when relate d to VWC dynam ic mainly cau sed by the seasonal variation (dry–w et cycle) of vegeta tion [26]. Seria l co ation can be used to forecast future values of the tim e-series by modeling the depe nd betwee n observ ations but affects correlati ons between variabl es measu red in time violates the bas ic reg ression assump tion of inde pendence [22]. Corre lation coefficie n seria lly correlated da ta cannot be used as ind icators of goodn ess-of -fit of a mo de correlati on co efficients are artificially infl ated [22,27 ]. The stu dy of the rela tionship betwee n NDWI and KBDI is a nontrivi al task due effect of ser ial correlati on. Rem edies for serial correlation inclu de sampl ing or aggr ing the data over longer tim e interva ls, as well as further mod eling, which can in techni ques such as weighted reg ression [25,28 ]. However , it is dif ficult to accou seria l co rrelation in time -series relate d to VWC dynam ics usin g exte nded regr techni ques. The time-serie s rela ted to VWC dyn amics often ex hibit hig h non-Gau seria l correlation and are more signifi cantly affec ted by outlie rs and measureme nt e [28]. A samp ling te chnique theref ore is pro posed, wh ich account s for seria l correl season al time-serie s, to stud y the rela tionship bet ween differe nt tim e-series. The correlati on effect in time-serie s is as sumed to be minimal wh en extracting one m per seaso n (e.g ., start of the dry season). The extracted seasonal metrics are then uti to study the relati onship betwee n time -series at a specif ic mome nt in tim e (e.g., the dry seaso n). The aim of this chapte r is to address the effect of seria l correlatio n wh en st the rela tionship betwee n remote sensing and mete orological time-se ries related to VW by comp aring nonse rially co rrelated season al metrics from time -series . Thi s ch theref ore has three defin ed obje ctives. Firstl y, an over view of time -series analysi s nique s and concepts (e.g ., stationar ity, auto correlatio n, ARIM A, etc.) is presente d the rela tionship betwee n tim e-series is studied usin g cros s-correla tion and or d least squa re (OLS) regres sion anal ysis. Secondly , an algorit hm fo r the extractio season al metrics is optimized for sa tellite and mete orological time -series. Final ly, tempo ral occu rrence and values of the extra cted nonse rially correlate d season al m are analyzed statistically to define the quantitative relationship between NDWI and KBDI time-series. The influence of serial correlation is illustrated by comparing results from cross-correlation and OLS analysis with the results from the investigation of correlation between extracted metrics. 8.2 Data 8.2.1 Study Area The Kruger National Park (KNP), located between latitudes 238S and 268S and longitudes 308E and 328E in the low-lying savanna of the northeastern part of South Africa, was selected for this study (Figure 8.1). Elevat ions ran ge from 26 0 to 839 m abov e se and mean annual rainfall varies between 350 mm in the north and 750mm in the south. The rainy season within the annual climatic season can be confined to the summer months (i.e., November to April), and over a longer period can be defined by alternating wet and dry seasons [7]. The KNP is characterized by an arid savanna dominated by thorny, fine-leafed trees of the families Mimosaceae and Burseraceae. An exception is the northern part of the KNP where the Mopane, a broad-leafed tree belonging to the Ceasalpinaceae, almost completely dominates the tree layer. ß 2007 by Taylor & Francis Group, LLC. Punda Maria N Shingwedzi Letaba Satara Pretoriuskop Onder Sabie 0 10 20 30 40 50 km FIGURE 8.1 The Kruger National Park (KNP) study area with the weather stations used in the analysis (right). South Africa is shown with the borders of the provinces and the study area (top left). 8.2.2 Climate Data Climate data from six weather stations in the KNP with similar vegetation types were used to estimate the daily KBDI (Figure 8.1). KBDI was derived from daily precipitation and maximum temperature data to estimate the net effect on the soil water balance [3]. Assumptions in the derivation of KBDI include a soil water capacity of approximately 20 cm and an exponential moisture loss from the soil reservoir. KBDI was initial-ized during periods of rainfall events (e.g., rainy season) that result in soils with maxi-mized field capacity and KBDI values of zero [8]. The preprocessing of KBDI was done using the method developed by Janis et al. [10]. Missing daily maximum temperat-ures were replaced with interpolated values of daily maximum temperatures, based on a linear interpolation function [30]. Missing daily precipitation, on the other hand, was assumed to be zero. A series of error logs were automatically generated to indicate missing precipitation values and associated estimated daily KBDI values. This was done because zeroing missing precipitation may lead to an increased fire potential bias in KBDI. The total percentage of missing data gaps in rainfall and temperature series was maximally 5% during the study period for each of the six weather stations. The daily KBDI time-series were transformed into 10-daily KBDI series, similar to the SPOT VEGETATION S10 dekads (i.e., 10-day periods), by taking the maximum of ß 2007 by Taylor & Francis Group, LLC. NDWI −KBDI 1999 2000 2001 2002 2003 FIGURE 8.2 The temporal relationship between NDWI ÿaKnBdDI time-series for the ‘‘Satara’’ weather station (Figure 8.1). each dekad. The negative of the KBDI time-se ries (i.e. , –KBD I) was analyzed in chapter such that the tempo ral dynam ics of KBDI and NDWI were related (Fi gure The –KBD I and ND WI are used through out this chapter . The Satara weathe r s central ly posi tioned in the st udy area, was sele cted to rep resent the tempo ral veg dynam ics. The other weath er statio ns in the study area demonstr ate similar temp vegeta tion dynam ics. 8.2.3 Remot e Sen sing Data The data set used is compo sed of 10-dai ly SPO T VEGE TATION (SPO T VGT) comp (S10 NDVI maximum value syn theses) acqui red over the stu dy area for the period 1998 to December 2002. SP OT VG T can provide local to global coverage on a re basis (e.g., daily for SPOT VGT). The synth eses result in sur face reflectan ce in the (0.43– 0.47mm), red (0.61– 0m.6m8 ), NIR (0.78– 0m.8m9 ), and SWI R (1.58–m1m.7)5 spect ral region s. Images were atmospher ically correc ted usin g the sim plified method for atm pheric correc tion (SM AC) [30]. The geome trically and radio metrica lly correc ted image s have a spat ial resolu tion of 1 km. The S10 SPOT VGT tim e-series were pre proce ssed to det ect data that erroneo influe nce the subseq uent fitting of function s to tim e-series, neces sary to def ine and metrics [6]. The imag e preproce ssing pro cedures per formed were: . Data points with a sa tellite view ing zenith angle (VZA8)wearbeovme a5sk0e d out as pixels locat ed at the very edge of the im> a5g0e.58)(VswZAath are affec ted by re-sam pling me thods that yield erroneo us spect ral val ues. . The ab errant SWIR detec tors of the SPO T VGT sensor, flag ged by the status m of the SPO T VGT S10 synthesis , also were masked out. . A data point was cl assified as cloud- free if the blue refle ctance was less than 0 [31]. The develop ed thr eshold appro ach was appl ied to identi fy cloud-free pixel for the study area. NDWI time-series were derived by selecting savanna pixels, based on the land cover map of South Africa [32], for a 3 3 pixel window centered at each of the meteorological ß 2007 by Taylor & Francis Group, LLC. ... - tailieumienphi.vn