Professional Documents
Culture Documents
net/publication/273313540
CITATIONS READS
4 353
2 authors:
Some of the authors of this publication are also working on these related projects:
Rapid assessment of Zimbabwe's water quality using GIS and Remote Sensing technologies View project
All content following this page was uploaded by Caleb Maguranyanga on 21 April 2016.
To cite this article: Caleb Maguranyanga & Amon Murwira (2014) Mapping Maize, Tobacco, and Soybean Fields in Large-Scale
Commercial Farms of Zimbabwe Based on Multitemporal NDVI Images in MAXENT, Canadian Journal of Remote Sensing: Journal
canadien de télédétection, 40:6, 396-405, DOI: 10.1080/07038992.2014.999914
Taylor & Francis makes every effort to ensure the accuracy of all the information (the “Content”) contained
in the publications on our platform. However, Taylor & Francis, our agents, and our licensors make no
representations or warranties whatsoever as to the accuracy, completeness, or suitability for any purpose of the
Content. Any opinions and views expressed in this publication are the opinions and views of the authors, and
are not the views of or endorsed by Taylor & Francis. The accuracy of the Content should not be relied upon and
should be independently verified with primary sources of information. Taylor and Francis shall not be liable for
any losses, actions, claims, proceedings, demands, costs, expenses, damages, and other liabilities whatsoever
or howsoever caused arising directly or indirectly in connection with, in relation to or arising out of the use of
the Content.
This article may be used for research, teaching, and private study purposes. Any substantial or systematic
reproduction, redistribution, reselling, loan, sub-licensing, systematic supply, or distribution in any
form to anyone is expressly forbidden. Terms & Conditions of access and use can be found at http://
www.tandfonline.com/page/terms-and-conditions
Canadian Journal of Remote Sensing, 40:396–405, 2014
Copyright
c CASI
ISSN: 0703-8992 print / 1712-7971 online
DOI: 10.1080/07038992.2014.999914
Abstract. In this study we used the maximum entropy (MAXENT) method to test the extent to which a combination of temporal
Downloaded by [University of Zimbabwe] at 08:12 16 January 2015
NDVI series and ancillary field data can be used to map the distribution of maize fields, soybean fields, and tobacco fields. We also
test whether there are specific times during the growing season suitable for separately mapping the fields. Results indicate that we
can separately map fields of maize, soybeans, and tobacco using remotely sensed data at specific times during the growing season.
We conclude that MAXENT can be used to map fields of maize, soybeans, and tobacco using a temporal series of NDVI data. In
addition, we identify critical periods that provide significant (p < 0.05) NDVI values useful for mapping.
Résumé. Dans cette étude, nous avons utilisé la méthode d’entropie maximale (MAXENT) pour évaluer dans quelle mesure une
combinaison de séries temporelles IVDN «Normalized Difference Vegetation Index; NDVI» et de données de terrain auxiliaires
peut être utilisée pour cartographier la répartition des champs de maı̈s, des champs de soja et des champs de tabac. Nous testons
également s’il existe des moments précis pendant la saison de croissance permettant de cartographier les champs séparément.
Les résultats indiquent que nous pouvons cartographier les champs de maı̈s, de soja et de tabac séparément à l’aide de données
de télédétection à des moments précis au cours de la saison de croissance. Nous concluons que MAXENT peut être utilisé pour
cartographier des champs de maı̈s, de soja et de tabac en utilisant une série temporelle de données NDVI. En outre, nous identifions
les périodes critiques qui fournissent des valeurs de NDVI significatives (p < 0,05) qui sont utiles pour la cartographie.
INTRODUCTION specific crops are less well developed, hence, they are a cur-
In Southern Africa, recurrent droughts and the threat posed rent focus of research. Knowledge of crop-specific statistics is
by a changing climate call for improved methods of estimat- critical for food security and crop-marketing decisions, thus,
ing crop production, especially in countries whose livelihoods it is critical to develop remote sensing approaches that can
are dependent on agriculture and are vulnerable to food short- provide crop-specific area statistics, as well as yield statistics
ages (Kandji et al. 2006; Ngigi 2009). One of the countries of any particular area (Kumar et al. 2012; Penigraphy et al.
highly dependent on agriculture in Southern Africa is Zim- 2009).
babwe, where about 70% of the population depends on agricul- Specific crop-area statistics derived from single remotely
ture for food, income, and employment (Government of Zim- sensed images taken at specific times of the growing season were
babwe 2010). Here, the objective of monitoring crop produc- one of the first approaches for mapping different crops based
tion patterns as a way to manage and estimate agricultural pro- on their spectral characteristics, particularly from the earliest
duction is critical for providing accurate information to policy Landsat satellites (Badhwar 1984; Rembold and Maselli 2006;
makers. Gusso and Ducati 2012). In recent years, the development of fine
The development of remote sensing and image processing temporal and moderate spatial resolution satellites such as the
techniques has allowed the mapping of large areas in a fast moderate resolution imaging spectroradiometer (MODIS) has
and economical way. Although remote sensing has been use- resulted in an increase of spatiotemporal-domain-based meth-
ful in mapping crop versus noncrop areas generally (DeFries ods to map specific crop types from remotely sensed data (Chang
and Townshend 1994; DeFries et al. 1998; Liu et al. 2005; et al. 2007). A study of time series data for individual crop types
Loveland et al. 2000; Xiao et al. 2005), methods for mapping in Kansas, USA, concluded that MODIS 250 Normalized Dif-
ference Vegetation Index (NDVI) data had sufficient spatial and
Received 3 April 2014; Accepted 26 November 2014. temporal resolution for major crop-type identification (Ward-
∗
Corresponding author e-mail:murwira@gis.uz.ac.zw; murwira low et al. 2007). The results of this work suggested that the spa-
@alumni.itc.nl tiotemporal domain is important for specific crop-area mapping.
396
VOL. 40, NO. 6, DECEMBER/DÉCEMBRE 2014 397
FIG. 2. Location of the digitized agricultural fields in the study area. Map in geographic coordinates based on the WGS 84
reference spheroid.
each field where observation was done after harvesting, follow- Remotely Sensed Crop Data
ing the methods of Sibanda and Murwira (2012). We used 71 For this study, we used the 16-day composite MODIS NDVI
samples for maize, 21 samples for soybean, and 15 samples for products (MOD13Q1), with a spatial resolution of 250 m,
tobacco with average field sizes of 22.03 ha, 20.89 ha and 16.41 covering the period between September 2010 and May 2011.
ha, respectively, as shown in Table 1. MODIS NDVI was used in this study because of its capability to
The samples fields are different, reflecting the proportions to provide information at an intermediate spatial resolution,
the area covered by each of these crops. The study area has more between a finer resolution of Landsat and coarser resolution
large-scale commercial fields of maize compared to soybean and of the advanced very high-resolution radiometer (AVHRR),
tobacco. Because the tobacco fields were large, we deemed the while still providing daily repeat coverage. In addition,
sampled area of tobacco adequate for modeling purposes. Landsat has a limited temporal resolution (up to 26 days)
that conceals many details of crop phenology dynamics. We
downloaded L3 Global 250 m MODIS 16-day data from
TABLE 1
Descriptive statistics of the maize, soybean, and TABLE 2
tobacco fields Descriptive statistics of the maize, soybean, and tobacco fields
using a confidence interval of 80%
Crop Mean Minimum Maximum Standard No. of
Type (ha) (ha) (ha) Deviation Pixels Cut off% (80) Crop Type Hectarage No. of Pixels
FIG. 3. NDVI profiles of maize, soybean, and tobacco for the 2010–2011 growing season. These results can be confirmed using
the Friedman test. The shaded areas show the periods that have significant NDVI values.
the USGS EROS Data Center2 (NASA 1999). These data tested whether there are any significant differences between crop
are produced by using a constrained maximum view angle NDVI among the image dates or among the different months,
approach (Huete et al. 2002). They were used because they are using the Friedman ANOVA for ranks. Furthermore, the Kendall
well calibrated and well validated to provide a substantially concordance coefficient was used to test whether the crops’ ND-
improved basis for scientific studies of this nature (Zhang VIs are similar (i.e., correlated). The Mann Whitney U test was
et al. 2003). The processing of these data includes atmospheric used to determine whether and at which time there was a sig-
correction, cloud removal, and bidirectional reflectance dis- nificant difference in the median NDVI values among maize,
tribution function (BDRF) correction (Chen et al. 2006). The soybean, and tobacco.
data were downloaded in the native integerized sinusoidal We then tested for the phenological stage at which we could
(ISIN) projection and reprojected to the geographic coordinate significantly distinguish among maize, soybean, and tobacco
projection system (lat/lon), World Geodetic System 1984 fields in a large-scale commercial farming area of Zimbabwe.
(WGS84) in a Geographic Information System (GIS). The This method was adopted from the one developed in the
data were then analyzed in the Integrated Land and Water Zambezi valley of Zimbabwe to distinguish cotton from maize
Information System Geographical Information System3 (ILWIS and sorghum in smallholder agricultural landscapes (Sibanda
GIS). and Murwira 2012). We used a temporal series of 16-day
The MODIS NDVI data obtained from the sampled sites were MODIS-derived NDVI data and GPS-surveyed crop-type
tested for spatial autocorrelation, using Moran’s I. We found no data of the 2010 to 2011 cropping season (September to
significant spatial autocorrelation in the data (p > 0.05). Next May).
we plotted the NDVI data graphically (Figure 3).
We then tested the data for any significant (p < 0.05) de-
viations from normality, using the Kolmogorov–Smirnov test
as a first step. Normality test results showed that data signifi- Modeling the Distribution of Maize, Soybean,
cantly (p < 0.05) deviated from a normal distribution. We also and Tobacco
We modeled the distribution of maize, soybean, and tobacco
2https://wist.echo.nasa.gov by using the MAXENT method (Phillips et al. 2004, 2006) and
2 environmental variables: NDVI and field boundaries. Specif-
3http://www.itc.nl ically we used the freely available MAXENT software version
400 CANADIAN JOURNAL OF REMOTE SENSING/JOURNAL CANADIEN DE TÉLÉDÉTECTION
Downloaded by [University of Zimbabwe] at 08:12 16 January 2015
FIG. 4. (a) Maize probability of occurrence. (b) Soybean probability of occurrence. (c) Tobacco probability of occurrence.
3.3.3k.4 MAXENT relies on presence-only occurrence records the jackknife approach (Phillips 2009). This approach excludes
to estimate the probability of occurrence for a species, which one variable at a time when the model is running. Consequently,
can then be used to discriminate suitable versus unsuitable ar- it provides information on the performance of each variable in
eas. MAXENT finds the probability distribution of maximum the model in terms of how important each variable is at explain-
entropy (i.e., that which is most spread out, or closest to uniform) ing the species distribution and how much unique information
and then constrains the distribution using a set of environmental each variable provides. In addition, several recent studies have
variables with a range of values defined by the environment at found the MAXENT model to perform well, or better, when
locations where the species is known to occur (Phillips et al. compared to other modeling methods (Kumar et al. 2009; Elith
2006). MAXENT is based on the premise that the unknown et al. 2006; Evangelista et al. 2008; Hernandez et al. 2006).
probability distribution should have maximum entropy but is MAXENT provides output data in raw, cumulative, and logis-
constrained by the environmental characteristics of the niche. tic formats (Phillips and Dudik 2008). Scores range from 0–1
MAXENT controls overfitting and variable selection using a (Phillips et al. 2006). The raw, cumulative. and logistic ouput
regularization that smoothes the modeled distribution, with a formats can be imported into a GIS to map probability distribu-
penalized maximum likelihood model that balances model fit tions. We used a 0.8 probability threshold to produce crop-type
with model complexity (Phillips and Dudik 2008; Elith et al. probability maps (maize, soybean, and tobacco) in ILWIS GIS.
2011). Furthermore, MAXENT determines how each variable Selection of the threshold was made following general rules
influences the presence of the modeled species in 2 ways. In the guiding the usefulness of the model based on scientific liter-
first approach, MAXENT provides the percent contribution of ature (Elith et al. 2006; Baldwin 2009; Peterson et al. 2007;
each variable to the final model (Baldwin and Bender 2008). The Pittman and Brown 2011; Araujo and Guisan 2006). We then
second alternative method for assessing variable importance is compared probabilities above 0.8. For each pixel, we retrieved
the crop with the highest probability to represent that crop type.
Maps showing the relative distribution of maize, soybean, and
4http://www.princeton.edu/∼schapire/maxent, accessed November 2011 tobacco were then produced.
VOL. 40, NO. 6, DECEMBER/DÉCEMBRE 2014 401
Downloaded by [University of Zimbabwe] at 08:12 16 January 2015
FIG. 5. The MAXENT model results showing the area under the receiver operating characteristic curve for (a) maize, (b) soybean,
and (c) tobacco.
Training the model, we used 70% of the training records, and 0.90 and 1 was considered to produce an excellent model. In
the remaining 30% were used to test the model performance. this study, based on exploratory data analysis, we adopted the
Validation and test data were randomly selected for each differ- Araujo and Guisan (2006) classification scheme.
ent crop-type grouping. Prediction by the model was achieved MAXENT calculates the contribution of each environmental
using the receiver operating curve (ROC). This curve is con- variable in building the model and its final output. Jackknife
structed by plotting the true positive rate (sensitivity) of the tests (Phillips et al. 2006) calculate the test, training, and AUC
model in the y-axis versus the false positive rate (1-specificity) gain resulting from each variable in the model on its own, in
in the x-axis. Sensitivity represents how well the data correctly its absence, and in the presence of all other variables. These
predicts presence, whereas specificity provides a measure of tests were used to determine which of the environmental vari-
correctly predicted absences (Baldwin 2009). The area under ables are most important in explaining the distribution of maize,
the ROC curve (AUC) is used as a measure of strength of the soybeans, and tobacco.
model or, in real terms, the probability that a cell chosen at
random will be classified correctly relative to a random sample
of background cells (Phillips et al. 2006). The possible values RESULTS
of the AUC fall between 0.5 and 1.0. The higher the AUC value, Figures 4a, 4b, and 4c illustrate the probability of occur-
the better the simulated results; and if the AUC value is less than rence of maize, soybean, and tobacco by location in the study
0.5, the simulated results were considered to be worse than ran- area based on the MAXENT model using digitized fields and
dom prediction (Fielding and Bell 1997). If the AUC > 0.5, the multitemporal NDVI variables for 70% observation data.The
model performed is considered better than random prediction probability of finding maize and soybean fields is high in Hu-
(Phillips et al. 2006), and if it is greater than 0.75, the prediction rungwe, Makonde, and Zvimba districts, whereas the prob-
is considered useful (Elith et al. 2011). Furthermore, following ability of finding tobacco fields is high in the Hurungwe
Araujo and Guisan (2006), if the AUC value of the model was district only. Furthermore, it can be observed that fields of
between 0.70 and 0.80, the model produced was considered av- high maize probability are close to fields with high soybean
erage; between 0.80 and 0.90 was good; and a value between probabilty.
402 CANADIAN JOURNAL OF REMOTE SENSING/JOURNAL CANADIEN DE TÉLÉDÉTECTION
Downloaded by [University of Zimbabwe] at 08:12 16 January 2015
FIG. 6. The MAXENT Model results showing the jackknife test of the importance of variables used in training the distribution of
(a) maize, (b) soybean, and (c) tobacco.
Using a confidence interval of 80%, we observe that maize the training data. The testing line indicates the fit of the model
has an area of 7543.75 hectares, soybeans an area of 1731.25 to the testing data, and is the real test of the model’s predictive
hectares, and tobacco an area of 1562.5. hectares. power. The diagonal straight line shows the line expected if the
Figure 5a illustrates the model evaluation using the ROC model performed no better than random (Phillips et al. 2006).
curve of the randomly selected training and test data for maize- We observe that the MAXENT model distinguished between
field prediction. The training line shows the “fit” of the model to probable and improbable maize fields because the AUC for the
VOL. 40, NO. 6, DECEMBER/DÉCEMBRE 2014 403
training data as well as for the test data is greater than that tobacco starts to mature and is the beginning of ripping; hence,
of the random prediction line (0.5). The AUC for the training the significantly different NDVI values.
data is 0.82 and for the test data is 0.81. In addition, the MAX- The difference between this study and other studies is in
ENT model distinguished between probable and improbable the explicit use of temporal NDVI data (in particular, pheno-
soybean fields. The AUC for the training data is 0.89 and for logical phases) and field boundary data in MAXENT to model
the test data is 0.85, which are all greater than the random pre- occurance of fields of maize, soybeans, and tobacco in a sin-
diction line (0.5; Figure 5b). Furthermore, the MAXENT model gle season. Other studies have emphasized climatic variables
distinguished well between probable and improbable tobacco in modeling crop distribution. Sun et al. (2012) revealed the
fields. The AUC for the training data is 0.77 and for the test data major climatic indices of the distribution of winter wheat culti-
is 0.76 (Figure 5c). vation zones, using MAXENT. In addition, Kemp (2012) used
Figure 6a illustrates that the environmental variable with the a MAXENT modeling technique to analyze information about
highest AUC gain when used in isolation is the NDVI for 6 environmental conditions at the current locations of agricultural
March 2011, whereas the environmental variable that decreases crops on Hawaii Island in order to predict the probability of suit-
the AUC gain most when used in isolation is the field boundaries able conditions existing for the same crops at other locations on
(here depicted as fieldsfinal). Figure 6b shows that the environ- the island. In particular, soil characteristics, rainfall, solar radi-
mental variable with the highest gain when used in isolation is ation elevation, slope, and temperature provided variables for
Downloaded by [University of Zimbabwe] at 08:12 16 January 2015
the NDVI for January 17, 2011, and the environmental variable the modeling effort. Among the results was the identification of
that decreases the most when used in isolation is fieldsfinal. various landscape conditions found in all locations where each
Figure 6c illustrates that the environmental variable with the indvidual crop is found.
highest gain when used in isolation is the NDVI for February Although temporal NDVI data analysis for the months under
2, 2011, and the environmental variable that decreases the most study perfomed well in our study area, they may not neces-
gain when used in isolation is fieldsfinal. sarily be the best candidates in other geographic regions or
spatial scales, thus, transfereability of the model may be lim-
ited. Further refinement of our methods is required, including
DISCUSSION tests that employ different sensors across multiple spatial scales.
Results of this study indicate that we can use a temporal Specifically, further research should consider precisely match-
NDVI series in MAXENT to separately map fields of maize, ing temporal NDVI data and spatial resolution. In addition, our
soybean, and tobacco at specific times during the growing sea- model results could have been improved if we integrated addi-
son. The high AUC for maize (0.81), soybean (0.85), and to- tional geospatial variables that characterize the physical land-
bacco (0.76) indicate the high predictive power of the MAXENT scape (e.g., soil characteristics, slope, rainfall, temperature, etc.)
model in mapping maize, soybean, and tobacco fields (Elith et al. taking cognizance of the need not to include large data layers in
2011). These results are supported by findings in Northeastern the analyses.
Thailand (Heumann et al. 2011) where MAXENT was suc-
cessfully used to model for upland cassava and lowland paddy CONCLUSION
rice varieties in a human-managed landscape. Furthermore, our The objective of this study was to determine the extent to
results are consistent with other findings that suggest the MAX- which a temporal series of remotely sensed vegetation indices
ENT algorithm is of particularly high utility for applications in (NDVI) and field boundaries used as environmental variables in
which presence data are limited (Evans et al. 2010; Phillips et al. MAXENT can map the distribution of maize, soybean, and to-
2006). bacco fields and identify the phenological stage (early greenup
With the MAXENT jackknife-based variable selection proce- onset, late greenup onset and green peak) that can best separate
dure, results indicate that we can identify critical days of NDVI maize, soybean, and tobacco within one growing season. We
variables, which are important for determining the separability conclude that we can separately map out fields of maize, soy-
of of maize, soybean, and tobacco fields. Specifically, results bean, and tobacco using remotely sensed data at specific times
indicate that the middle of January and early March (midgrow- during the growing season. Specifically, we confirm that the
ing season) correspond with the critical periods of late greenup middle of January 2011 and the early of March 2011, which
onset and green peak that provide significant (p < 0.05) NDVI correspond with the late greenup onset and green peak are the
values for mapping maize, soybean, and tobacco fields. This is critical periods that provide signifcant (p < 0.05) NDVI values
consistent with the crop calendars of the 3 crops. Specifically, for mapping maize, soybean, and tobacco fields.
soybean crops are planted about 2 weeks after maize, however,
during the period January 6 to March 6, 2011 (late greenup REFERENCES
phase to green peak) soybean has higher NDVI values than Araujo, M. B., and Guisan, A. 2006. “Six (or so) research priorities for
maize, whereas, earlier in the growing season, maize has higher species distribution modeling.” Journal of Biogeography, Vol. 33:
NDVI values compared with soybean. This is also the time when pp. 1677–1688.
404 CANADIAN JOURNAL OF REMOTE SENSING/JOURNAL CANADIEN DE TÉLÉDÉTECTION
Badhwar, G. D. 1984. “Automatic corn-soybean classification using Heumann, B. W., Walsh, S. J., and McDaniel, P. M. 2011. “As-
Landsat MSS data, I, near-harvest crop proportion estimation.” Re- sessing the application of a geographic presence-only model for
mote Sensing of the Environment, Vol. 14: pp. 15–29. land suitability mapping.” Ecological Informatics, Vol. 6(No. 5):
Baldwin, R. A. 2009. “Use of maximum entropy modeling pp. 257–269.
in wildlife research.” Entropy, Vol.11(No. 4): pp. 854–866, Huete, A., Didan, K., Miura, T., Rodriguez, E. P., Gao, X., and Fer-
doi:10.3390/e11040854. reira, L. G. 2002. “Overview of the radiometric and biophysical
Baldwin, R. B., and Bender, L. C. 2008. “Den-site characteristics of performance of the MODIS vegetation indices.” Remote Sensing of
black bears in Rocky Mountain National Park, Colorado.” Journal Environment, Vol. 83(No. 1–2): pp. 195–213.
of Wildlife Management, Vol. 72: pp. 1717–1724. Kandji, S. T., Verchot, L., and Mackensen, J. 2006. Climate Change
Chang, J., Hansen, M. C., Pittman, K., Carroll, M., and DiMiceli, Climate and Variability in Southern Africa: Impacts and Adaptation
C. 2007. “Corn and soybean mapping in the United States using in the Agricultural Sector. World Agroforestry Centre (ICRAF),
MODIS time-series data sets". Agronomy Journal, Vol. 99(No. 6): United Nations Environment Programme (UNEP).
pp. 1654–1664. Kemp, K. K. 2012. The Hawai’i Island Crop Probability Map. Los
Chen, P., Fedosejevs, G., Tiscareno-Lopez, M., and Arnold, J. 2006. Angeles, CA, USA: Spatial Sciences Institute.
“Assessment of MODIS-EVI, MODIS-NDVI and Vegetation-NDVI Kumar, A., Roy, P. S., Ghosh, S. K., and Gilbert, I. 2012. “Effect
composite data using agricultural measurements: an example of at on specific crop mapping using WorldView-2 multispectral add-
corn fields in Mexico.” Environmental Monitoring and Assessment, on bands: soft classification approach.” Journal of Applied Remote
Downloaded by [University of Zimbabwe] at 08:12 16 January 2015
Phillips, S. J., Dudik, M., and Schapire, R. E. 2004. “A maximum Sun, J.-S., Zhou, G.-S., and Sui, X.-H. 2012. “Climatic suitability of the
entropy approach to species distribution modeling.” In Proceedings distribution of the winter wheat cultivation zone in China.” European
of the Twenty-First International Conference on Machine Learning, Journal of Agronomy, Vol. 43: pp. 77–86.
Banf, Canada, July 2004. Vincent, V., and Thomas, R. G. 1960. An agricultural survey of South-
Phillips, S. J., and Dudik, M. 2008. “Modeling of species distribu- ern Rhodesia. Part 1: agroecological survey. Salisbury, UK: Gov-
tions with Maxent: new extensions and a comprehensive evaluation.” ernment Printer.
Ecography, Vol. 31: pp. 161–175. Wardlow, B. D., Egbert, S. L., and Kastens, J. H. 2007. “Analysis of
Pittman, S. J., and Brown, K. A. 2011. “Multi-scale approach for pre- time-series MODIS 250 m vegetation index data for crop classifica-
dicting fish species distributions across coral reef seascapes.” PLoS tion in the U.S. Central Great Plains.” Remote Sensing of Environ-
ONE, Vol. 6(No. 5): e20583. doi:10.1371/journal.pone.0020583. ment, Vol. 108(No. 3): pp. 290–310.
Rembold, F., and Maselli, F. 2006. “Estimation of inter-annual crop Xiao, X., Boles, S., Liu, J., Zhuang, D., Frolking, S., and Li, C. 2005.
area variation by the application of spectral angle mapping to low “Mapping paddy rice agriculture in southern China using multi-
resolution multitemporal ndvi images.” Photogrammetric Engineer- temporal MODIS images.” Remote Sensing of Environment, Vol.
ing and Remote Sensing, Vol. 72(No. 1): pp. 55–62. 95(No. 4): pp. 480–492.
Sibanda, M., and Murwira, A. 2012. “The use of multi-temporal Zhang, X., Friedl, C. B., Schaaf, A. H., Strahler, J. C. F., Hodges,
MODIS images with ground data to distinguish cotton from maize F., Gao, B. C., et al. 2003. “Monitoring vegetation phenology
and sorghum fields in smallholder agricultural landscapes.” Interna- using MODIS.” Remote Sensing of Environment, Vol. 84(No. 3):
Downloaded by [University of Zimbabwe] at 08:12 16 January 2015
tional Journal of Remote Sensing, Vol. 33(No. 16): pp. 4841–4855. pp. 471–475.