TEMPORAL RELATIONSHIPS BETWEEN REMOTELY SENSED SOIL MOISTURE AND NDVI
OVER AFRICA: POTENTIAL FOR DROUGHT EARLY WARNING?
ZIPPORAH MUSYIMI March, 2011
SUPERVISORS:
Dr. A, Vrieling Dr. M, Schlerf
Thesis submitted to the Faculty of Geo-Information Science and Earth Observation of the University of Twenty in partial fulfilment of the requirements for the degree of Master of Science in Geo-information Science and Earth Observation.
Specialization: Natural resource management
SUPERVISORS:
Dr. A, Vrieling Dr. M, Schlerf
THESIS ASSESSMENT BOARD:
Prof. W. Verhoef (Chair)
Prof. T. Udelhoven (External Examiner)
TEMPORAL RELATIONSHIPS BETWEEN
REMOTELY SENSED SOIL MOISTURE AND NDVI OVER AFRICA: POTENTIAL FOR DROUGHT EARLY WARNING?
ZIPPORAH MUSYIMI
Enschede, The Netherlands, March, 2011
DISCLAIMER
This document describes work undertaken as part of a programme of study at the Faculty of Geo-Information Science and Earth Observation of the University of Twente. All views and opinions expressed therein remain the sole responsibility of the author, and do not necessarily represent those of the Faculty.
The temporal relationship between remotely sensed near surface soil moisture and Normalised Difference Vegetation Index time series (2003-2009) for the African continent has been analysed using correlation analysis and distributed lag (DL) models. For the correlation analysis, optimal lag was the lag that reflects the best NDVI-soil moisture correlation. Up to 4 optimal lags were identified for the correlation analysis.
The lags were analysed with vegetation formation data, soil data and climate data. Optimal lag 0 was dominant in open and closed forest. Optimal lag 1 was more dominant in grasslands, croplands and in shrub lands. For the DL models, in most areas significant time delayed relationships between soil moisture and NDVI was confined up to lag 5. For the DL analysis, optimal lag was the lag that best reflects the lagged responses of NDVI anomalies on soil moisture anomalies. Up to 4 optimal lags were identified for the DL analysis. The results of the optimal lags were analysed with vegetation formation, soil and climatic data. Optimal lag 0 was dominant in open and closed forest in areas with Greysols and Vertisols. Lag 1 was most dominant in croplands especially in steppe and savannah climatic conditions, while lag 2 was most dominant in grasslands and in very arid (desert) climatic conditions. Generally, areas indicating NDVI dependence on lagged values of soil moisture were confined to grasslands and croplands in arid areas and semi and semi-arid areas.
This work would not have been accomplished without dedicated support from my supervisors Dr. Anton Vrieling and Dr. Martin Schlerf. Dr. Vrielig you ensured that I had all that was needed to accomplish this thesis. The long hours of discussions we had trying to figure out the direction that this work would take are highly appreciated. Your technical support, guidance and thoughts are highly appreciated. Also most appreciated is your rigorous editing on my manuscripts, which you never gave up on. It was really nice working with you and learnt a lot from you. Martin your door was always open for me. You helped me consolidate my scattered ideas into a compact and precise document and even when I had to submit my work at the last minute, you never gave up on me and always read my work and guided me all through.
Thanks for your support, and input towards this work. Anton and Martin I could not have desired for a better supervision. I also appreciate the support and guidance for Professor Thomas Udelhoven from the university of Trier. The discussions we had helped me have a clear mind on how to approach my thesis. I also thank you for your advice thereafter during my data analysis
I Thanks ITC for admitting me as a student following my application to this institution. My sincere thanks go to NUFFIC for sponsoring my education.
To my friends here in ITC and back home thanks a lot for your moral support. Rachna. Thanks for your support, understanding and friendship.
Mum and dad you are the best parent’s one would ever ask for. Thanks for your support and encouragement. You have always have faith me and encouraged me when I was almost giving up. To all my brothers thanks your moral support.
.
Abstract ... i
Acknowledgements ... ii
List of figures ... v
List of tables ... vi
1. Introduction ... 1
1.1. Background ... 1
1.2. Objectives ... 3
1.3. Research questions ... 3
1.4. Hypothesis ... 3
1.5. Thesis structure ... 3
2. Materials and Methods ... 5
2.1. Description of the data sets ... 5
2.2. Description of the methods ... 7
2.2.1. Exploratory data analysis ... 7
2.2.2. Evaluation of the NDVI and soil moisture correlation ... 9
3. Results ... 15
3.1. Spatial variability of NDVI and soil moisture ... 15
3.2. NDVI – soil moisture correlation ... 17
3.3. Temporal dependence of NDVI anomalies on soil moisture anomalies ... 26
4. Discussion ... 33
5. Conclusion ... 37
6. List of references ... 39
is illustrated in black, and the filtered profile is illustrated in red. After filtering results into a smooth series.
... 6
Figure 2: Decomposed NDVI time series of a point at a location (3°15'11.79"N, 43°49'30.39"E)in East Africa. The observed was decomposed to trend, seasonal and the random component. The NDVI values presented here are scaled to 8 bits ... 11
Figure 3: Decomposed soil moisture time series of a point at a location (3°15'11.79"N, 43°49'30.39"E) in ... 12
Figure 4: (a) shows the original series (b) detrended series at location 3°15'11.79"N, 43°49'30.39"E in East Africa ... 12
Figure 5: Mean NDVI and soil moisture for the period between 2003-09. The white part represents areas without data either for the soil moisture or for NDVI ... 15
Figure 6 NDVI time plots indicating the temporal behaviour at Point 1, 2, 3 and 4 during the observation period (2003-2009). The NDVI values represent DN values. ... 17
Figure 7 : NDVI time plots indicating the temporal behaviour for Point 1, 2, 3 and 4 during the observation period ... 18
Figure 8: Scatter plot of NDVI versus soil moisture at different lags at point 1. ... 19
Figure 9: Scatter plot of NDVI versus soil moisture at different lags at point 2. ... 19
Figure 10: Scatter plot of NDVI versus soil moisture at different lags at point 4. ... 20
Figure 11: NDVI – Soil moisture correlation structure with different lags ... 21
Figure 12: NDVI –soil moisture optimal lags and their corresponding correlation coefficient ... 23
Figure 13: Illustration of spatial coverage of the soil types in areas with a correlation coefficient(r) of >0.7 and a sensitivity of >50%. The white parts indicates areas with no valid data for the soil moisture and the NDVI series or areas with r <0.7 or in areas with a soil type with a sensitivity of <50%. ... 25
Figure 16: DL model results illustrating the significance of the soil moisture regression coefficient at concurrent (zero lag) up to lag 5(50 days shift) ... 27
Figure 17: DL optimal lags with their corresponding optimal correlation coefficient (r2 )for the NDVI – soil moisture relationship. ... 29
Table 1: Summary description of the main specifications of the soil moisture and NDVI data set ... 6 Table 2: Lists of the original GlobCover land cover classification and the main vegetation formations derived from aggregating different vegetation covers ... 8 Table 3: Major soil types used in this study adapted from the FAO UNESCO soil type classification. Level 1 classification corresponds to the major soil type consists of aggregation of all level two classes ... 9 Table 4: Locations of points whose temporal characteristics of both the soil moisture and NDVI are discussed. ... 16 Table 5: Evaluation results of optimal lags identified from the correlation analysis of soil moisture and NDVI with vegetation formations. Sensitive areas are areas with an optimal correlation coefficient (r)>
0.7. Insensitive areas are which had an r < 0.7. Sensitive and insensitive areas are expressed as a percentage of the total area covered by a given soil type. The total area covered by each lag is expressed as a
percentage of the sensitive area for a given vegetation formation type. Only soil types which had a
sensitivity of > 50% were considered. ... 21 Table 6: Evaluation results of optimal lags identified from the correlation analysis of soil moisture and NDVI with soil types. Sensitive areas are areas with an optimal correlation coefficient (r)> 0.7. Insensitive areas are which had an r < 0.7. Sensitive and insensitive areas are expressed as a percentage of the total area covered by a given soil type. The total area covered by each lag is expressed as a percentage of the sensitive area for a given soil type. Only soil types which had a sensitivity of > 50% were considered. ... 24 Table 7: Evaluation results of optimal lags identified from the correlation analysis of soil moisture and NDVI climate types. Sensitive areas are areas with an optimal correlation coefficient (r)> 0.7. Insensitive areas are which had an r < 0.7. Sensitive and insensitive areas are expressed as a percentage of the total area covered by a given climate type. The total area covered by each lag is expressed as a percentage of the sensitive area for a given climate type. ... 26 Table 8: Evaluation results of the DL optimal lags based on vegetation formations. Sensitive areas are areas where optimal lags were identified and had a correlation coefficient (r2 ) > 0.5. Insensitive areas are areas where with non-significant relationship or areas with a r < 0.5. For each vegetation formation, total area covered by each lag is expressed as a percentage of the sensitive area. ... 30 Table 9:Results of evaluating the DL optimal lags based on soil types. Sensitive areas are areas where optimal lags were identified and had a correlation coefficient (r2 ) > 0.5. Insensitive areas are areas where with non-significant relationship or areas with r < 0.5. The sensitive and insensitive are expressed as a percentage of the total area covered by a given soil type. For each soil type the total area covered by each lag in the sensitive areas is expressed as a percentage of the total sensitive area of a given soil type ... 30 Table 10: Evaluation results of the DL optimal lags based on climate types. Sensitive areas are areas where optimal lags were identified and had a correlation coefficient (r2) > 0.5. Insensitive areas are areas where with non-significant relationship or areas with a r < 0.5. For each climate type, total area covered by each lag is expressed as a percentage of the total area covered by the sensitive area for a given climate. ... 31 Table 11: Summary results of evaluation of the DL optimal lags with vegetation formation type, soil type and climate data. Only vegetation formations, soils and climate types with a sensitivity of >70% were considered. All the vegetation, soil and climate data in which a given optimal lag was dominant area grouped together. ... 31
1. INTRODUCTION
1.1. Background
Drought is a major recurrent problem in many parts of Africa (Wilhite, 2000). Since a large part of African agriculture is rain fed (Anyamba, et al., 2005), and mostly in water limited areas (Unganai, et al., 2005) drought events are commonly related to reduction in crop production and in severe cases total crop failure (FAO, 2010). A huge proportion of the African population directly depend on agriculture for livelihood and sustainability, and hence drought events are commonly associated hunger, famines, malnutrition and starvations (FAO, 2010). Climate change predictions predict increase in drought frequency of occurrence, intensity and persistence in Africa in the future (Boko, et al., 2007). As such drought monitoring and early warning remains of prime importance in the African (World Bank, 2005). Drought being a natural phenomenon there is nothing much a society can do to prevent its occurrence (Wilhite, 2000). However, early warned of the likelihood of a drought event, the risks associated with droughts can be minimised (Nicholson, et al., 1990; Unganai, et al., 2005). For example, early warned, governments and international food aid agencies can initiate strategic plans in advance such as internal redistribution of food, food aid requests and strategic food aid missions (Sannier, et al., 1998) to mitigate food insecurity in drought venerable areas.
The ability of satellite remote sensing to provide data in a spatially continuous manner and on regular basis makes it the most important data source for drought monitoring (Boken, 2005). Most common used remotely sensed data in drought monitoring in Africa include long term Normalised Difference vegetation Index (NDVI) (Rouse, et al., 1974) and Rainfall Estimates (RFEs) (Rowland, et al., 2005). The main operational institutions involved in drought monitoring and early warning such in Africa such as Drought Monitoring Centre in Nairobi (DMCN) and SADC Regional Early Warning System (REWS), Famine Early Warning Systems Network (FEWSNET) (http://www.fews.net/Pages/default.aspx?l=en), Global Information and Early Warning System (http://www.fao.org/giews/english/index.htm) (http://www.fao.org/giews/english/index.htm) and the Monitoring of Agriculture Resource (http://mars.jrc.ec.europa.eu/mars) use either historical and near real time NDVI or RFEs or both.
These institutions use these data sets to identify anomalous conditions as indicators for drought (Rowland, et al., 2005). FEWSNET, SADC REWS and FAO GIEWS use course resolution NDVI derived from Advanced Very High Resolution Radiation (AVHRR) (Mukhala, 2005; Rowland, et al., 2005) while MARS FOODSEC unit uses NDVI images from SPOT VEGETATION onboard SPOT 5 platform (http://www.spot-vegetation.com/). FEWSNET and GIEWS use RFEs images to analyse rainfall variability from the historical data. FEWSNET uses REF images National Oceanic and Atmospheric Administration (NOAA) climate production centre REF data while GIEWS REFs from Cold Cloud estimation (CCD) from the meteorological satellite (METEOSAT) (Rowland, et al., 2005).
Monitoring vegetation dynamics using satellite remote sensing is commonly achieved by utilising the strong coupling between the visible (Owe, et al.) and the NIR reflectance with the physiological condition of the leaves and their density (Anyamba, et al., 2005). Over the years, the red and NIR radiances have been converted to different vegetation indices, which are used as proxies for estimating various vegetation
used NOAA-AVHRR NDVI, Tucker, et al. 1985, used NOAA-AVHRR for estimation of total herbaceous biomass production in the Senegalese Sahel, Tucker, et al. 1986 utilised NOAA-AVHRR NDVI for monitoring grassland in the Sahel in 1984-85. Also several studies have been used to investigate the relationship between NDVI and climatic variations such as rainfall (Herrmann, et al., 2005; Heumann, et al., 2007; Nicholson, 2000; Nicholson, et al., 1990; Richard, et al., 1998). Though these studies differ in detail notably the temporal aggregation of NDVI (Foody, 2003), they all indicate a strong temporal dependence of NDVI on climatic variables in particular rainfall. It is on the basis of this strong relationship between rainfall, that NDVI is commonly used to indicate rainfall in areas where it is hard to get accurate rainfall data (Foody, 2003).
Soil moisture is the ideal parameter to monitor in drought monitoring (Boken, 2005) as drought results to soil moisture deficiency below the wilting point (Wilhite, 2000). Since Droughts are normally characterised by large spatial coverage, their monitoring would require data sets with large spatial coverage as well as comparable unbiased long-term observations. Unfortunately, accurate soil moisture estimates for large spatial scale been difficult to make. This is because soil moisture is highly heterogeneous spatially and temporally (de Jeu, et al., 2008; Njoku, et al., 2003) and hence very difficult parameter to measure not at point observation but for a large spatial coverage (Jackson, et al., 1989; Tilman, et al., 2001). As an alternative, soil moisture surrogates such as precipitation are commonly used in drought monitoring (Boken, 2005). However, application of soil moisture surrogates to infer drought conditions may fail to give a real reflection of the drought conditions. In the recent years with advancements in satellite technology, research has shown that global soil moisture can be monitored from operational satellite platforms particularly from satellite microwave (De Jeu, et al., 2003; Njoku, et al., 2005; Owe, et al., 2008;
Owe, et al., 2001; Wagner, et al., 1999). Evaluation of the correspondence of these satellites remotely sensed soil moisture products with ground measurements have shown promising results (Draper, et al., 2009; McCabe, et al., 2005; Rüdiger, et al., 2009). The promising results of soil moisture estimates from microwave remotely sensed soil moisture has opened more research opportunities to evaluate the potential utility of these products in agricultural applications (Bolten, et al., 2010).
Distributed Lag model DL is a special kind of a regression, which accounts for the lagged time responses between variables (Dominic, et al., 2002). This method was used to investigate the relationship between rainfall and NDVI in east Africa (Eklundh, 1998). Ji et al., (2005) used the same method to investigate the lag and seasonality responses of NDVI to precipitation in Central Great Plain in US. Udelhoven et al., (2009) used distributed lag model to assess the relationship between NDVI and rainfall anomalies in Spain.
This study aims at evaluating the temporal relationships between satellite retrieved soil moisture from Advanced Microwave Scanning Radiometer (AMSR-E) and vegetation dynamics for the African continent. This evaluation is essential in identifying areas within the continent where vegetation dynamics exhibits a strong temporal dependence on soil moisture. In such areas, analysis of AMSR-E soil moisture could have practical application in drought monitoring and early warning.
1.2. Objectives General objective
The general objective of this study is to assess the temporal relationships between remote sensed soil moisture (AMSR-E) and NDVI in Africa with an aim of identifying areas where this soil moisture product can be used as an early indicator for drought for early warning.
Specific objective
The specific objectives of this study are:
To spatially evaluate the relationships between soil moisture and NDVI dynamics for the African continent
Assess the temporal dependency of NDVI anomalies on soil moisture anomalies
1.3. Research questions
How is soil moisture –NDVI correlation and can an optimal time lag be identified for different areas?
Do vegetation formation types, soil types and climatic conditions influence the correlation between soil moisture and NDVI?
Is the relationship between NDVI and soil moisture anomalies significant at 90% confidence level?
Is the relationship between NDVI and soil moisture anomalies influenced by vegetation formations, soil type and climatic conditions?
1.4. Hypothesis
For different areas optimal time lags can be identified from the soil moisture –NDVI correlation
The correlation between soil moisture and NDVI is influenced by vegetation formation, soil type and climatic conditions
The relationship between NDVI and soil moisture anomalies is significant at 90% confidence level
The relationship between NDVI and soil moisture anomalies is influenced by vegetation formation, soil type and climatic conditions
1.5. Thesis structure
Chapter 1 consists of background information on the study, objectives, research questions hypothesis and outline of the thesis structure. Description of the data sets and methods used is provided in chapter 2.
Results and interpretation of the results are provided in chapter 3. The results are discussed in chapter 4.
Chapter 5 provides a brief summary of the study and conclusions and finally the references.
2. MATERIALS AND METHODS
2.1. Description of the data sets
This study utilised two main data sets derived from remotely sensed data, namely soil moisture and NDVI.
Below these two data sets are described in detail. To evaluate the results, additional data sets were used.
These additional data sets are described in brief.
Soil moisture data
The soil moisture estimates derived from Advanced Microwave Scanning Radiometer of the Earth Observing System (AMSR-E) was used. This soil moisture product is one of the longest and consistent existing passive microwave soil moisture data. Operational specifications of AMSR-E are well documented (Li, et al., 2003; Njoku, et al., 2003; Njoku, et al., 2005). AMSR-E signals can penetrate only the top soil profile (1-2 cm) and hence the soil moisture estimates relates to the near surface soil moisture (de Jeu, et al., 2008). Only the descending (night pass) recording of the sensor were used. This is because during the night thermally, the surface is homogenous than during the day thus allowing for better parameter approximation (Owe, et al., 2008). Among the existing soil estimates from AMSR-E (Jackson, 1993;
Koike, et al., 2004; Njoku, et al., 2003; Owe, et al., 2001), the soil moisture estimate retrieved using the algorithm developed by Vrije Universiteit Amsterdam (VUA) in collaboration with NASA commonly known as VUA-NASA was used (Owe, et al., 2001). The VUA-NASA soil moisture estimate was prefered over the other products for two reasons. First, evaluation studies have shown that the VUA-NASA results outperforms other algorithms showing a high correspondence with in situ data (Wagner, et al., 2007;
Draper, et al., 2009) and modeled soil moisture (Rüdiger, et al., 2009). Secondly, among all the existing AMSR-E soil moisture products, the VUA-NASA product is extensively validated (de Jeu, et al., 2008).
The VUA-NASA product is retrieved using the Land Parameter Retrieval Model (LPRM) (De Jeu, et al., 2003; Holmes, et al., 2009; Owe, et al., 2008; Owe, et al., 2001). LPRM links surface geophysical variables (i.e. soil moisture, vegetation water content, and soil/canopy temperature) to the observed brightness temperatures (de Jeu, et al., 2008) through forward radiative transfer modeling (de Jeu, et al., 2003). The main specifications of the soil moisture data are shown in Table 1.
The daily temporal resolution of the soil moisture dataset was downscaled to the 10-daily NDVI resolution by compositing the daily products to 10 days. As daily data does not have complete coverage (the AMSR-E descending swath allows to cover the tropics once every two days), the temporal compositing also resulted in a continuous spatial product. Compositing was done in IDL by calculating for each pixel the median of 10 daily products to obtain a representative soil moisture value for the 10-day period. To remove remaining noise from the 10 day soil moisture series, the series were filtered with the Savitzky-Golay algorithm (Savitzky, et al., 1964) using a fourth order polynomial. Figure 1 depicts that filtering results in a smoother series.
Table 1: Summary description of the main specifications of the soil moisture and NDVI data set
Figure 1: Soil moisture profiles at a location in Chad (3°15'11.79"N, 43°49'30.39"E). The unfiltered profile is illustrated in black, and the filtered profile is illustrated in red. After filtering results into a smooth series.
NDVI data set
SPOT Vegetation (SPOT VGT) NDVI dataset (http://www.vgt.vito.be/) was used. The NDVI data is derived from the red and the NIR bands and provided in DN values ranging from 0 to 255. The data sets contains 10- day maximum value NDVI composites at 1 km resolution from July 2002 to December2009.
The 10-day composites are generated by VITO using Maximum Value Composites (MVC) algorithm.
MVC aims at reducing the effects of atmospheric interferences such as clouds. The main characteristics of the NDVI data set are illustrated Table 1. NDVI values are strongly influenced by atmospheric conditions Characteristics Data sets
Soil moisture NDVI
Sensor EMSR-E SPOT VGT
Source (http://geoservices.falw.vu.nl/adaguc_portal _dev/)
http://free.vgt.vito.be/.
Spatial resolution 0.25 degrees 1Km
Temporal resolution Daily 10 days
Temporal coverage 1st Jan 2003 to 31st December 2009 1st 2003 to 31st December 2009
Projection WGS84 WGS84
Units Volumetric values (m3 m-3). DN values (1-255)
such as clouds and haze. Although maximum value composting substantially reduces these effects (Chen, et al., 2004), atmospheric residuals still persist as noise. To eliminate this residual noise from the datasets, the iterative Savitzky-Golay algorithm (Savitzky, et al., 1964) as described by Chen et al (2004) was applied.
To jointly evaluate NDVI with the soil moisture dataset, the 1km NDVI data set was downscaled to 0.25 degrees resolution. Because the precise resolution of the NDVI data-set was defined as 0089286 degrees, precisely 28 x 28 pixels were contained in one 0.25 degree resolution cell. Down scaling was achieved by averaging (for each 10-day period) the NDVI values within the 28 by 28 pixels. To exclude water affected pixels from the NDVI aggregates, water bodies were masked with water bodies shape file from the FAO GEONetwork (http://www.fao.org/geonetwork/srv/en/metadata.show?id, #154). If 50% or more of the output pixel was water then no results were calculated. Conversely, if less than 50% of the out put pixel was water then aggregate value was calculated as the average of the land pixels only.
Land cover data
To investigate whether the relationship between soil moisture and NDVI is influenced by vegetation formations, vegetation formations were derived from the 2007 GlobCover V2.2 land cover classification.
Compared to existing other global land cover datasets (De Fries, et al., 1998; De Fries, et al., 1994;
Loveland, et al., 1999), GlobCover V2.2 has a finer resolution of 300m, has been extensively validated, and its 22 land cover classes are compatible with those of the United Nations Food and Agriculture Organisation’s (FAO) Land Cover Classification System (LCCS). Based on FAO land cover classes, five vegetation formations namely closed forests, open forests, shrub lands, grasslands and croplands as shown in Table 2
Soil data
To investigate whether the relationship between soil moisture and NDVI is influenced by soil type, the soil data from FAO-UNESCO Soil Map of the World was used. This soil map consists of digitised soil types at a scale of 1:5.000.000 in vector format. The soil types were aggregated into major soil types as indicated in Table 3.
2.2. Description of the methods
Analysis undertaken in this study can be categorised into three. First, prior to indepth analysis, exploratory data analysis were performed to get a general understanding of the data. Second Pearson correlation analysis was performed to evaluate the relationship between soil moisture and NDVI. Lastly, a regression analysis was performed using NDVI and soil moisture anomalies to establish a link between soil moisture and NDVI. These analysis and how their results were evaluated are explained below.
2.2.1. Exploratory data analysis
The mean NDVI and soil moisture during the observation period were visually analysed to evaluate their correspondence i.e if areas with high mean NDVI corresponds well with areas with high mean soil moisture and vice versa. Prior to complex time series analysis it is strongly recommended to visualize the
Table 2: Lists of the original GlobCover land cover classification and the main vegetation formations derived from aggregating different vegetation covers
(Adapted from GlobCover V2.2 land cover classification)
time series observation against time (Shumway, et al., 2006). Such plots are called time plots and play a critical role in understanding the structure of the series (Metcalfe, et al., 2009) as well as exposing missing data or outliers (Chartfield, 2004). As such the soil moisture and NDVI series over time were visualised in time plots.
GlobCover land cover classification Land cover classes used
in this study Post-flooding or irrigated croplands (or aquatic)
Crop lands Rain fed croplands
Mosaic cropland (50-70%) / vegetation (grassland/shrub land/forest) (20- 50%)
Mosaic vegetation (grassland/shrub land/forest) (50-70%) / cropland (20- 50%)
Grasslands Closed to open (>15%) grassland or woody vegetation on regularly
flooded or waterlogged soil - Fresh, brackish or saline water
Closed to open (>15%) herbaceous vegetation (grassland, savannas or lichens/mosses)
Mosaic forest or shrub land (50-70%) / grassland (20-50%)
Shrub lands Mosaic grassland (20-50%) / forest or shrub land (70-80%)
Sparse (<15%) vegetation
Closed (>40%) broadleaved deciduous forest (>5m)
Closed Forest Closed (>40%) needle leaved evergreen forest (>5m)
Closed to open (>15%) broadleaved forest regularly flooded (semi- permanently or temporarily) - Fresh or brackish water
Closed (>40%) broadleaved forest or shrub land permanently flooded - Saline or brackish water
Closed to open (>15%) (broadleaved or needle leaved, evergreen or deciduous) shrub land (<5m)
Open (15-40%) broadleaved deciduous forest/woodland (>5m)
Open Forest Open (15-40%) needle leaved deciduous or evergreen forest (>5m)
Closed to open (>15%) mixed broadleaved and needle leaved forest (>5m)
Table 3: Major soil types used in this study adapted from the FAO UNESCO soil type classification. Level 1 classification corresponds to the major soil type consists of aggregation of all level two classes
FAO- UNESCO soil type codes Major soil type
Level 1 code Level 2 code
G Gc, Gd, Ge, Gh, Gm, Gp, Gx, Gleysols
Q Qa, Qc, Qf, Ql, Arenosols
O Od, Oe, Ox Histosols
L La, Lc, Lf, Lg, Lk, Lo, Lp, Lv Luvisols
N Nd, Ne, Nh, Nitosols
R Rc, Rd, Re, Rx Regosols
Z Sg, Zm, Zo, Zt Solonchaks
A Af,Ag, Ah, Ao, Ap Acrisols
T Th, Tm, To, Tv Andosols
B Bc, Bd, Be, Bf, Bg, Bh, Bk, Bv, Bx, Cambisols
F Fa, Fh, Fo, Fp, Fr, Fx Ferralsols
J Jc, Hd, Je, Jt, Fluvisols
K Kh, Kk, Kl, Kastaznozems
I Lithosols
H Hc, Hg, Hh, Hl, Phaeozems
W Wd, We, Wh, Wm, Wx, Planosols
P Pf, Pg, Ph, Pl, Po, Pp, Podisols
E Rendzinas
S Sg, Sm, So, Solonetz
V Vertisosl
X Xh, Xk, Xl, Xy, Xerosols
Y Yh, Yk, Yl, Yt, Yy Yermosols
(Adapted from FAO UNESCO )
2.2.2. Evaluation of the NDVI and soil moisture correlation
Prior to correlation analysis, NDVI series at a given pixel were plotted against soil moisture series in scatter plots. There is always a time lag in which vegetation growth takes to responds to soil moisture (Foody, 2003). Thus in each of the scatter plots, NDVI values at a time t are plotted against values of soil moisture at a time t and its lagged values of up to lag 8. This implies that NDVI values at a time t is plotted against soil moisture values at a time t-0, up to t-8.
Pearson correlation on pixel basis was performed to evaluate the strength of the relationship between soil moisture and NDVI. Previous studies investigating the relationship between NDVI and climatic variables especially rainfall have indicated a lag response between peak NDVI precipitation events (Eklundh, 1998;
Ji, et al., 2005; Li, et al., 2002; Nicholson, et al., 1990). As such, the NDVI was also correlated with lagged variables of soil moisture. Lag 0 implied that soil moisture values at time t was compared to NDVI values at time t while lag 1 indicated that soil moisture at time t-1 was correlated to NDVI at time t.
Previous studies investigating the relationship between NDVI and climatic variables in Africa indicate that the relationship is influenced by vegetation (Nicholson, et al., 1990) and soil types (Nicholson, et al., 1994) and climatic conditions. Thus, the optimal lags were evaluated with soil, vegetation formation and climatic data. For the vegetation formation, each type of vegetation formation was categorised as sensitive and insensitive. Sensitive areas were defined as those areas with a correlation coefficient (r) of more than 0.7, which corresponds to an r2 of approximately 0.5. On the other hand, insensitive areas were defined as those areas with a correlation coefficient of less than 0.7. The total areas of each lag in the sensitive areas of each vegetation formation were computed to evaluate the relationship between the lags and different vegetation and soil and climate type. The same method was repeated using climate data and soil type data.
evaluate the presence of trends and seasonal variations both the soil moisture and NDVI series were decomposed. Trend and seasonal variations were then eliminated from the series. Lastly regression analysis was performed. These methods are step wise elaborated below.
Decomposition of the time series
Both the soil moisture and NDVI series were decomposed in R statistical computing software. The series were decomposed into the observed, trend, seasonal and error components using simple additive decomposition model of the form (Metcalfe, et al., 2009):
xt = mt + st + zt (Equation 1) where at time t, xt is the observed series, mt is the trend, st is the seasonal effect and zt is the an error term which is a sequence of stationary points. Figure 2 and Figure 3 illustrate the decomposed NDVI and soil moisture series respectively at a location 3°15'11.79"N, 43°49'30.39"E in East Africa. The composed series was just to visualise trends and seasonal variations and was not used in any analysis.
100140180
observed 125135145
trend -2002040
seasonal -30-101030
2003 2004 2005 2006 2007 2008 2009 2010
random
Time
Decomposition of additive time series
Figure 2: Decomposed NDVI time series of a point at a location (3°15'11.79"N, 43°49'30.39"E)in East Africa. The observed was decomposed to trend, seasonal and the random component. The NDVI values presented here are scaled to 8 bits
Elimination of trends and seasonal effect
The decomposed time series depicts that both the NDVI (Figure 2) and soil moisture (Figure 3) series were influenced by trends. To assure stationary conditions of the series the non linear trend in both the series was removed using Fourier polynomials in TimestatsV1.0. using a base period of 252 dekads (the length of the total series). Figure 4 illustrates the original NDVI series and the series after detrending. The decomposed time series also depicts that both the NDVI and soil moisture series were influenced by the seasonal variations. Seasonal variation strongly influences the autocorrelation structure of a time series (Udelhoven, et al., 2009). Performing multivariate analysis with autocorrelated data lead to spurious results (Chartfield, 2004). The seasonal effect was eliminated by calculating the seasonal anomalies. This method was elsewhere by Udelhoven et al (2009) while assesing the relationship between environmental factors (temperature and rainfall) in Spain.
Decomposed NDVI series
304050
observed 4042444648
trend -8-40246
seasonal -6-2246
2003 2004 2005 2006 2007 2008 2009 2010
random
Time
Decomposition of additive time series
Figure 3: Decomposed soil moisture time series of a point at a location (3°15'11.79"N, 43°49'30.39"E) in East Africa. The observed is decomposed to trend, seasonal and the random component.
Years
NDVI
2003 2004 2005 2006 2007 2008 2009 2010
100140180
Years
NDVI
2003 2004 2005 2006 2007 2008 2009 2010
80120160200
Figure 4: (a) shows the original series (b) detrended series at location 3°15'11.79"N, 43°49'30.39"E in East Africa
Decomposed soil moisture series
Original NDVI series NDVI series without trends component
(a) (b)
The seasonal anomalies were calculated from the detrended series in TimestatsV1.0. according to Udelhoven et al., 2009
seas
t X
X
d (Equation 2)
where Xt is the variable at a time t and x seas is the long term seasonal mean at time t.
To reduce variance, the seasonal anomalies were standardised using Z score standardization:
j j tj
tj S
X Z X
(Equation 3) where
Xj is the long term seasonal mean, Sj is the standard deviation, and j and t is the time index and Xtj is the season at a time t
Distributed lag model
After eliminating trends and seasonal variations, a regression analysis was performed using the NDVI and soil moisture seasonal anomalies using DL models. DL regression analysis accounts for the lagged effects of explanatory variable on the response variable and takes the form :(Dominic, et al., 2002)
t t
t t
t a b x b x b x u
y 0 1 1 2 2 ... (Equation 5)
Vector b describes the weight assigned to past x series and u is the model residual.
Considering the soil moisture as the explanatory variable and NDVI as the response variable the above equation takes the form:
t j t i
i j
t a b soil moisture u
NDVI
max 0
(Equation 6)
where, bj = the impulse response weight vectors describing the weights assigned to current and past soil
parameter estimation is achieved by the iterative generalised least square approach to enable consistent and efficient parameter estimation for auto correlated series.
Optimal lags from the DL analysis were computed on pixel based. An optimal lag for DL was defined as the lag whose t statistics were significant (both negative and positive) and had the lowest p value at 90%
confidence level when the p values of all significant lags were considered. The DL optimal lags results were analysed with vegetation formations soil and climatic data
3. RESULTS
3.1. Spatial variability of NDVI and soil moisture
Figure 5 illustrates spatial variability of the mean 10-day NDVI and soil moisture during the study period.
The NDVI values are not the real (-1to +1) NDVI values but have been were scaled to 8 bit (0 to 255). In the whole document reference to NDVI refers to the scaled values. In the transition zone between the Sahara desert and the equatorial zone, NDVI clearly increases from the north to the south. Mean soil moisture shows a similar North-South gradient in this region.
Figure 5: Mean NDVI and soil moisture for the period between 2003-09. The white part represents areas without data either for the soil moisture or for NDVI
In Eastern Africa, in most areas the spatial variability of the mean NDVI corresponds well with means soil moisture. For example, arid regions in Northern Somalia depict low values of both the soil moisture and NDVI. Areas around Lake Victoria and the Ethiopian highlands depict high values for both the soil
50°0'0"E 50°0'0"E
30°0'0"E 30°0'0"E
10°0'0"E 10°0'0"E
0°0'0"
0°0'0"
40°0'0"N 30°0'0"N
20°0'0"N 10°0'0"N
0°0'0" 0°0'0"
10°0'0"S 10°0'0"S
30°0'0"S 30°0'0"S
<VALUE>
No data 10-80 81-150 151-230
±
Mean NDVI (2003-2009)
50°0'0"E 50°0'0"E
30°0'0"E 30°0'0"E
10°0'0"E 10°0'0"E
0°0'0"
0°0'0"
40°0'0"N20°0'0"N0°0'0"10°0'0"S30°0'0"S
0 7501,500 3,000 4,500 Km
±
Mean soil moisture (2003-2009)
Africa1
<VALUE>
No data 0 - 13 14- 25 25 - 65
contrarily to the soil moisture, which indicates persistence of low values north of the Kalahari Desert, NDVI indicate an increasing pattern.
The temporal characteristics of the two series were visually inspected in time plots. Table 4 gives locations of points whose timeplots for both the soil moisture and NDVI are are illustrated and discussed. Point 1 is at alocation in the transition zone between the Sahara deseert and the equitorial zone. Point 2 is at a location in East Africa. Point 3 is at a location in the southern Africa region. Point 4 is at a location in the East coast of Magadascar.
Table 4: Locations of points whose temporal characteristics of both the soil moisture and NDVI are discussed.
Point Lat Lon Region
1 3°15'11.79"N 43°49'30.39"E Transition zone between Sahara and equatorial zone and the Sahel region
2 11°55'49.54"N 18°12'52.21"E East Africa
3 17°23'24.57"S 17°27'32.65"E Southern Africa
4 18°28'57.71"S 48°55'52.81"E Madagascar
Figure 6 and Figure 7 illustrates time plots of NDVI and soil moisture respectively at locations listed in table 4. The NDVI temporal dynamics of the point 1, 2 and 3 are comparable to that of the soil moisture.
However, the soil moisture and NDVI temporal characteristics of point 4 differ significantly. This location is characterised by high rainfall though out the year. Thus, at this location, soil moisture is not a limiting factor for the vegetation development and NDVI temporal dynamics are insensitive to soil moisture dynamics.
Years
NDVI
2003 2004 2005 2006 2007 2008 2009 2010
80120160
Years
NDVI
2003 2004 2005 2006 2007 2008 2009 2010
100140180
Years
NDVI
2003 2004 2005 2006 2007 2008 2009 2010
80120160200
Years
NDVI
2003 2004 2005 2006 2007 2008 2009 2010
210220230240
Figure 6 NDVI time plots indicating the temporal behaviour at Point 1, 2, 3 and 4 during the observation period (2003-2009). The NDVI values represent DN values.
3.2. NDVI – soil moisture correlation
Figure 8, Figure 9 and Figure 10 illustrates NDVI and soil moisture scatter plots matrix for point 1, 2 and 4 respectively. Figure 8 depicts positive relationship between NDVI and soil moisture at lag 0, lag 1 and lag 2. From lag 3 onwards though still positively related, there is a lot of scatter which progressively increases with increasing lags. Figure 9 indicates a weak positive relationship and the relationship weakens as the lags increase. Figure 10 indicates almost no relationship between the two signals for all lags.
Point 1 Point 2
Point3 Point 4
Years
Soil moisture
2003 2004 2005 2006 2007 2008 2009 2010
05102030
Years
Soil moisture
2003 2004 2005 2006 2007 2008 2009 2010
303540455055
Years
Soil moisture
2003 2004 2005 2006 2007 2008 2009 2010
101520253035
Years
Soil moisture
2003 2004 2005 2006 2007 2008 2009 2010
45505560
Figure 7: Soil moisture time plots indicating the temporal behaviour for Point 1, 2, 3 and 4 during the observation period
Optimal lags and correlations
Figure 11 gives a visualization of the NDVI –soil moisture correlation by illustrating the optimal lags and their corresponding optimal correlation coefficients on a pixel basis. Figure 12 gives a summarised visualization of optimal lags and their corresponding correlation coefficient with an r>0.7. The optimal lags were mainly constrained in lag 0, 1, 2 and 3 which corresponds to the correlation of NDVI at a time t
=0 with lagged values of soil moisture at a time t=0, t=1 t=2 and t=3. From lag 5 onwards, the correlation coefficient decreased for all pixels.
Point 1 Point 2
Point 3 Point 4
Figure 8: Scatter plot of NDVI versus soil moisture at different lags at point 1.
Figure 10: Scatter plot of NDVI versus soil moisture at different lags at point 4.
Optimal lag 0 was dominant in the southern parts of the Sahel strip covering areas in the Southern Sudan, Chad, and Nigeria; many parts of Togo; Northern parts of Ghana, Mali, and some parts of Mali and Guinea. Optimal lag 1 was most wide spread in East Africa covering areas in eastern and North-eastern parts of Kenya. In many parts of Southern Africa Lag 1 was the most dominant. Considering only optimum correlation coefficient of above 0.7, the transition zone between the Sahel and the equatorial zone was dominated by high correlation coefficient (r) of between 0.9 to 0.99. In Eastern Africa, optimal correlation coefficients between 0.8 and 0.89 were the most dominant. In the Southern African region, optimal correlation coefficients of between 0.8,- 0.89, and 0.9 - 1 were most dominant.
Table 5 illustrates result of evaluating the influence of vegetation formations on the NDVI -soil moisture correlation. Among the five vegetation formations, grasslands were the most sensitive while closed forests were the least sensitive. Croplands had a moderate sensitivity of 66 %. For croplands, grasslands and shrub lands, most of the sensitive areas were covered by optimal lag 1 and 2. In contrast, for the closed forest and the open forests the sensitive areas were mostly dominated by lag 0.
50°0'0"E 50°0'0"E
30°0'0"E 30°0'0"E
10°0'0"E 10°0'0"E
10°0'0"W 10°0'0"W
40°0'0"N 40°0'0"N
20°0'0"N 20°0'0"N
0°0'0" 0°0'0"
20°0'0"S 20°0'0"S
0 1,450 2,900 5,800 8,700
Km 1:90,386,104
Projection:GCS_W84 Datum: 1984 Musyimi 2011
50°0'0"E 50°0'0"E
30°0'0"E 30°0'0"E
10°0'0"E 10°0'0"E
10°0'0"W 10°0'0"W
40°0'0"N20°0'0"N0°0'0"20°0'0"S
Optimal lags for NDVI soil moisture correlation
Optimal correlation coefficient for NDVI-soil moisture correlation
Legend Lags
Lag 4 Lag 3 Lag 2 Lag 1 Lag 0
Legend Coefficient (r)
0.7-0.79 0.8-0.89 0.9-0.99
´ ´
Figure 11: NDVI – Soil moisture correlation structure with different lags
Table 5: Evaluation results of optimal lags identified from the correlation analysis of soil moisture and NDVI with vegetation formations. Sensitive areas are areas with an optimal correlation coefficient (r)>
0.7. Insensitive areas are which had an r < 0.7. Sensitive and insensitive areas are expressed as a percentage of the total area covered by a given soil type. The total area covered by each lag is expressed as a percentage of the sensitive area for a given vegetation formation type.
Croplands (Area %)
Grasslands (Area%)
Shrublands (Area %)
Open forests (Area%)
Closed forest (Area%)
Insensitive 34 24 45 66 81
Sensitive 66 76 55 54 19
Optimal lag 0 16 14 24 55 67
Optimal lag 1 54 46 38 29 22
Optimal lag 2 31 37 22 10 8
Optimal lag 3 5 3 12 4 2
Optimal lag 4 2 3 3 2 1
TEMPORAL RELATIONSHIPS BETWEEN REMOTELY SENSED SOIL MOISTURE AND NDVI OVER AFRICA: POTENTIAL FOR DROUGHT EARLY WARNING? Figure 12: NDVI –soil moisture optimal lags and their corresponding correlation coefficient. The white areas indicate areas with no data or areas r<0.7
Projection:GCS_W84 Datum: 1984 Musyimi 2011
Lag 0Lag 1 Lag 2 Lag 3 Lag 4
´
coefficients 0.7-0.79 0.8-0.89 0.9-0.99Coefficients 0.7-0.79 0.8-0.89 0.9-0.99
Coefficients 0.7-0.79 0.8-0.89 0.9-0.99 Coefficients 0.7-0.79 0.8-0.89 0.9-0.99
Coefficients 0.7-0.79 0.8-0.89 0.9-0.99
02,7005,4008,1001,350 Km
Optimal lags and their corresponding correlation coefficients 1:136,767,725
