Investigating the Use of High Resolution Multi-spectral Satellite Imagery

for Crop Mapping in Nigeria

Crop and Landuse Classification using WorldView-3 High Resolution

Multispectral Imagery and LANDSAT8 Data

Tunrayo Alabi, Michael Haertel

and Sarah Chiejile

Geospatial Laboratory, International Instititute of Tropical Agriculture (IITA),

PMB 5320, Oyo Road, 200001 Ibadan, Nigeria

Keywords: Crop Mapping, WorldView-3, LANDSAT8, Cassava, Maize, Nigeria, Land Cover Classification,

Maximum Likelihood, Neural Network, Support Vector Machine.

Abstract: Imagery from recently launched high spatial resolution WorldView-3 offers new opportunities for crop

identification and landcover assessment. Multispectral WorldView-3 at 1.6m spatial resolution and

LANDSAT8 images covering an extent of 100Km² in humid ecology of Nigeria were used for crop and

landcover identification. Three supervised classification techniques (maximum likelihood(MLC), Neural

Net clasifier(NNC) and support vector machine(SVM)) were used to classify WorldView-3 and

LANDSAT8 into four crop classes and seven non-crop classes. For accuracy assessment, kappa coefficient,

producer and user accuracies were used to evaluate the performance of all three supervised classifiers. NNC

performed best with an overall accuracy(OA) of 92.20, kappa coefficient(KC) of 0.83 in landcover

identification using WorldView-3. This was closely followed by SVM with an OA of 91.77%, KC of 0.83.

MLC performed slightly lower at an OA of 91.25% and KC of 0.82. Classification of crops and landcover

with LANDSAT8 was best with MLC classifier with an OA of 92.12% , KC of 0.89. Cassava at younger

than 3 months old could not be identified correctly by all classifiers using WorldView-3 and LANDSAT8

products. In summary WorldView-3 and LANDSAT8 data had satisfactory performance in identifying

different crop and landcover types though at varying degrees of accuracies.

1 INTRODUCTION

Agriculture is crucial to man’s livelihood as the

major source of food. Feeding the growing human

population which is expected to reach more than 9

billion by 2050 could pose a serious challenge in the

midst of the uncertainties and complexities of the

predicted future climate. There will be need to

constantly boost agriculture production in a

sustainable and efficient way (Foley et al., 2011). To

achieve this, dependable, accurate and

comprehensive agricultural intelligence on crop

production is imperative. Agricultural production

monitoring can support decision-making and

prioritization efforts towards ameliorating

vulnerable parts of agricultural systems. The value

of satellite Earth observation data in agricultural

monitoring is well recognized (Low and Duveiller,

2014) and a variety of methods have been developed

in the last decades to provide agricultural production

related statistics (Carfagna and Gallego, 2005)

Remotely sensed data from satellite platforms

such as LANDSAT and SPOT have been used to

inventory a wide variety of earth resources,

including agricultural land and crops. The synoptic

overview provided by these satellite systems at

regular intervals has allowed farmers and

agricultural scientists to obtain information

concerning the condition of crops grown over a large

area. Satellite imagery has been used for crop

species identification and area estimation since

1970s. Much research has been carried out in the use

of LANDSAT MSS and TM data to estimate and

identify crops. Various authors have found out that

within some reasonable limits of accuracy, crops can

be identified in LANDSAT MSS, TM/ETM (Xavier

et al., 2005; Yang et al, 2007) or SPOT imagery

(Hanna et al., 2004; Xavier et al., 2005).

Alabi, T., Haertel, M. and Chiejile, S.

Investigating the Use of High Resolution Multi-spectral Satellite Imagery for Crop Mapping in Nigeria - Crop and Landuse Classiﬁcation using WorldView-3 High Resolution Multispectral

Imagery and LANDSAT8 Data.

In Proceedings of the 2nd International Conference on Geographical Information Systems Theory, Applications and Management (GISTAM 2016), pages 109-120

ISBN: 978-989-758-188-5

109

Many researchers have reported the use of multi-

temporal imagery within a given year to map

agricultural crops (Brewster et al., 1999) which has

tremendous advantages. However, in tropical

environment, cloud cover can limit this approach.

Classifying remotely sensed data remains a

challenge because many factors, such as the

complexity of the landscape in a study area, selected

remotely sensed data, and image-processing and

classification approaches, may affect the success of

a classification. Major limitations on crop

identification with satellite imagery relate to the

similarity of plant reflectance of different crops in

the available spectral bands, field-to-field variability

of plant reflectance of the same crops, the particular

combination of crops grown in a given region, the

pattern of individual crop phenology, spatial and

spectral variability within fields (Buechel et al.,

1989; Vassilev, 2013; Yang et al., 2007).

Moreover agricultural field in Africa are often

small in size and very often many different plant

species are found in a very small area (always the

case if they are intercropped) which makes the

homogeneous crop identification process rather

difficult with coarse resolution satellite imagery

(Campbell, 1996).

Advancements in digital image processing and

geographic information systems (GIS) have

increased the potential for deriving more accurate

crop information from satellite imagery (Ehrlich et

al., 1994; Rodrıguez et al., 2006).

High resolution satellite imagery offers more

opportunity in crop identification. From 1999 when

IKONOS was launched, several other High

resolution satellites such as Quick Bird,

WorldVIEW 1, 2 and 3 or Pleaides followed. The

competition between these multi spectral platforms

led to decreasing prices per km² with resolutions up

to 30 cm per pixel.

Various researchers have evaluated the use of

these modern satellite products for land cover types,

crop classification in diverse regions of the world

(Ozdarici-Ok et al., 2015; Srestasathiern and

Rakwatin, 2014; Yang et al., 2007) but the use of

these products for crop identification in humid

regions like Nigeria and other tropical areas in West

Africa has not been well documented.

Hence the objective of this study is to evaluate

satellite images captured by the newly launched

WorldView-3 sensor for crop identification and land

cover classification in Nigeria as well as the use of

LANDSAT8 OLI for landcover and cropland

mapping.

2 STUDY AREA

The study location is the Ore Agricultural Village in

Ondo state, Nigeria. The Agriculture village is

dedicated to crop farming and animal husbandry and

is situated on a 3000-hectare facility. The Ondo

State Agricultural Village at Ore was created and

started operation in 2011 as a tool for empowering

the youth, the women and adults through agriculture

and represents one of three integrated Agricultural

villages established in the state that have been

established in order to reduce unemployment among

the younger population.

Participants at the village are drawn from young

graduates who have just completed their Higher

National Diploma and Bachelor degree and who are

willing to take up agriculture as a career. 

The natural vegetation of the site is tropical

rainforest characterized by pockets of secondary

forest and fallow regrowth. The area is characterized

by a length of growing period of more than 270 days

with humid forest ecology. The annual mean

maximum temperature at the site is 31.5 °C while

the minimum is 22.1°C. Mean annual rainfall is

about 2067 mm. Rainfall starts around March and

continues till middle of November. The topography

of the land varies from nearly flat to moderately high

slope. Mean elevation in the farm land is about 132

m above sea level with a mean slope of 8.7%.

Nearly 25% of the study area has slope greater than

12%. Major soil type of the farm is Ferric Lixisols

(Sonneveld, 2005). Soil texture is coarse loamy

sand, imperfectly or poorly drained.

During the late season of 2014, Cassava and

Maize were the major crops planted on the fields

within the target area. Cassava (Manihot esculenta)

is a perennial woody shrub with an edible root which

grows in tropical and subtropical areas of the world.

Cassava is the third largest source of food

carbohydrates in the tropics, after rice and maize

(Fauquet and Fargette, 1990). Cassava is a major

staple food in the developing world, providing a

basic diet for over half a billion people (It is one of

the most drought-tolerant crops, capable of growing

on marginal soils. Nigeria is the world's largest

producer of cassava, while Thailand is the largest

exporter of dried cassava.

In 2014, cassava and maize were planted during

the late season of August through November.

Specifically, the first batch of cassava was planted

on September 10

and the second was planted on

November 13

at end of rainy season. The total area

of cassava planted was 219 Ha. Apart from the

young cassava planted, matured cassava plots of

GISTAM 2016 - 2nd International Conference on Geographical Information Systems Theory, Applications and Management

110

between 12-15 months old were also found in the

study area, often mixed with weeds. These matured

cassava farms with weeds represent typical plots in

West Africa. Farmers stop weeding their cassava

field once they reach 5-6 months. During ground

truth field visit several of such plots have been

observed from which a few of them have been

selected to serve as a training site for the

classification process to be able to classify such

ready for harvest cassava fields.

Maize was planted for late season from August

through September 30

. The total area of maize

was 100 ha. Other non-crop land cover types were

also classified. Such includes primary forest,

degraded forest, roads, and rivers, mixed

fallow/shrubby grassland and bare ground.

Figure 1: Study area showing WorldView-3 natural colour

image acquired January 3rd, 2015.

3 SATELLITE IMAGERY

3.1 WorldView-3

WorldView-3 was launched on 13th August, 2014 in

California. It is the first multi-payload, super-

spectral, high-resolution commercial satellite

featuring 16 multispectral bands (eight in visible and

NIR spectrum and eight in the SWIR spectrum).

Operating at an altitude of 617 km, WorldView-3

provides 31 cm panchromatic resolution, 1.24 m

multispectral resolution, 3.7 m short-wave infrared

resolution, and 30 m CAVIS resolution.

WorldView-3 has an average revisit time of <1 day

(1m GSD) and is capable of collecting up to 680,000

km- per day.

A new tasking order for WorldView-3 image

was placed at around October 2014 covering an

extent of 100 km² which is the minimum extent for a

tasking order from Digital Globe Inc. (Longmont,

Colorado). Only the first eight multispectral bands

of the WorldView-3 were purchased. The SWIR

bands were not available for purchase at the time of

order. Due to much cloud cover in the region, we

could only obtain cloud free image on January 3,

2015. The geographic coordinates at the center of

the area are (Longitude 4.7922047°E, Latitude

6.731706° N). The spatial resolution of the imagery

was 1.24 m and the dynamic range of the data was

16 bits. Prior to delivery, the imagery was

radiometrically and geometrically corrected and

rectified to the world geodetic survey 1984

(WGS84) datum and the universal transverse

Mercator (UTM) coordinate system of Zone 31N.

Figure 2: Classification results of WorldView-3 (a) Maximum Likelihood, (b) Neural Net and (c) Support Vector Machine.

Investigating the Use of High Resolution Multi-spectral Satellite Imagery for Crop Mapping in Nigeria - Crop and Landuse Classiﬁcation

using WorldView-3 High Resolution Multispectral Imagery and LANDSAT8 Data

111

Table 1: Specifications of eight multispectral and panchromatic bands of World-View 3 sensor.

Specifications Multispectral sensor Panchromatic sensor

Spatial resolution 1.24 (m) 40 cm

Radiometry 16 bits 16 bits

Spectral bands

1. Coastal Blue (400 to 450 nm)

2. Blue (450 to 510 nm)

3. Green (510 to 580 nm)

4. Yellow (585 to 625 nm)

5. Red (630 to 690 nm)

6. Red-Edge (705 to 745 nm)

7. NIR1 (770 to 895 nm)

8. NIR2 (860 to 1040 nm)

(450 to 800 nm)



3.2 LANDSAT8 Data

LANDSAT program of the United States of America

is the longest running enterprise for acquisition of

satellite imagery of Earth. On July 23, 1972 “the

Earth Resources Technology Satellite” was

launched. This was eventually renamed to

LANDSAT. The most recent, LANDSAT8 was

launched on February 11, 2013 which provided

continuity in LANDSAT earth observation mission

(Lulla et al., 2013). The LANDSAT8 orbits our

planet every 99 min, covering the entire earth every

16 days except for the highest polar latitudes.

LANDSAT8 follows a sun-synchronous orbit at an

average altitude of 705 km and 98.2° inclination (Jia

et al., 2014)

The data quality (signal-to-noise ratio) and

radiometric quantization (12-bits) of the

LANDSAT8 Operational Land Imager (OLI) and

Thermal Infrared Sensor (TIRS) are higher than

previous LANDSAT instruments (8-bit for TM and

ETM+). The OLI sensor aboard LANDSAT8 has

nine bands for capturing the spectral response of the

earth's surface at discrete wavelengths along the

electromagnetic spectrum. Additionally, the TIRS

sensor aboard LANDSAT8 collects information at

two discrete wavelengths within the thermal infrared

portion of the electromagnetic spectrum. These

wavelengths have been chosen carefully based on

years of scientific research. For the study area, cloud

free LANDSAT8 images with path/row: 190/55

acquired on December 14, 2014 and January 15,

2015 was downloaded from the “earth explorer”

website (http://earthexplorer.usgs.gov/).

4 METHODS OF IMAGE

CLASSIFICATION

As an image analysis software EXELIS ENVI 5.2

was used to classify both satellite images. Many

different supervised classification techniques are

available in ENVI 5.2 including minimum distance,

Mahalanobis distance, maximum likelihood, neural

networks, and support vector machine. Maximum

likelihood is probably the most commonly used

classifier even though other classifiers may offer

advantages for some applications. In this study the

following algorithms are explored: Maximum

likelihood (MLC), Support Vector Machine (SVM)

and Neural Network (NNC) due to their high

performance reported in literature (Foody & Mather,

2004, Pal and Mather, 2005, Omkar et al, 2008).

The Maximum Likelihood Classifier (MLC) is

a well-known parametric statistical classifier and is

widely used for pattern classification (Duda and

Hart., 1973). A normal distribution is assumed for

the input data which include two parameters - mean

vectors and covariance matrices of the class

distributions are estimated and used in the

discriminant functions. MLC is generally accepted

as a standard against which the performance of other

classification algorithms is compared with (Omkar

et al, 2008).

Support Vector Machine (SVM) is a supervised

classification method derived from statistical

learning theory that often yields good classification

results from complex and noisy data. It separates the

classes with a decision surface that maximizes the

margin between the classes. The surface is often

called the optimal hyperplane, and the data points

closest to the hyperplane are called support vectors.

The support vectors are the critical elements of the

training set (Vapnik, 1979; Zhu and Blumberg,

2002). SVM can be adapted to become a nonlinear

classifier through the use of nonlinear kernels. While

SVM is a binary classifier in its simplest form, it can

function as a multiclass classifier by combining

several binary SVM classifiers (creating a binary

classifier for each possible pair of classes). ENVI’s

implementation of SVM uses the pairwise

GISTAM 2016 - 2nd International Conference on Geographical Information Systems Theory, Applications and Management

112

classification strategy for multiclass classification.

SVM has been shown to also work well for crop

classification (Foody and Mather, 2004; Pal and

Mather, 2005; Jia et al., 2014)

Artificial Neural Networks (ANNs) were

originally designed as pattern-recognition and data

analysis tools that mimic the neural storage and

analytical operations of the brain. ANN approaches

have a distinct advantage over statistical

classification methods in that they are non-

parametric and require little or no a priori

knowledge of the distribution model of input data.

Additional superior advantages of ANNs include

parallel computation, the ability to estimate the non-

linear relationship between the input data and

desired outputs, and fast generalization capability.

Many previous studies on the classification of

multispectral images have confirmed that ANNs

perform better than traditional classification methods

in terms of classification accuracy, such as

maximum likelihood classifiers (Yuan et al. 2009).

More detailed discussion on ANNs can be found in

Lippmann, 1987 and Richards and Jia 2006. In a

recent work Sandoval et al. 2014 used ANNs to

perform crop classification and obtained satisfactory

results.

4.1 Supervised Classification

Ground truth field visit was conducted on 4

February, 2015 from which georeferenced photos of

each crop and landcover type were collected. These

photos were used for developing training sites for

each crop or cover type class. Eleven cover types

were identified: Matured maize class consisted of

maize at full maturity by 3 January 2015 when

WorldView-3 image was taken, though a few of

them would still maintain some green colour by

December 14, 2014 when the first LANDSAT8 data

was acquired. Young cassava class are 3 month old

cassava by January 3, 2015 when WorldView-3

image was taken. Very young cassava class

consisted of cassava planted in the first and second

week of November, 2014 and were just one and half

month old by January 2015. This class consists more

of bare ground. Matured cassava class consisted of

those that were observed on the field planted over a

year before January 2015. This class shows typical

matured cassava fields mixed with weeds, shrubs

and trees. Other land cover types identified on the

land are: Degraded Forest, Primary Forest,

Fallow/grassland, Built up, Major River, Bare

ground/dirt road and Tarred road.

Supervised training sites were created using

online digitizing in ArcGIS 10.3 on known crops or

cover types with the aid of ground truth geotagged

photos.

Training samples were created proportional in

size and number to each land cover type extent.

Supervised classification using Maximum

likelihood algorithms in ENVI 5.2 with default

parameters; probability threshold set to none and

data scale factor of 1 was used to classify the eight

multispectral bands of WorldView-3 and

LANDSAT8 OLI into the eleven classes. Coastal

blue band was removed initially to see whether its

exclusion will improve accuracy of classification,

but it was found that using all bands gave a slightly

higher accuracy using confusion matrix tool

accuracy assessment. Hence eight bands were used

for WorldView-3. A similar procedure was followed

for LANDSAT8 classification. Coastal aerosol band

1 and Cirrus band 9 were removed from supervised

classification after the method of Jia et al., 2014, but

it was found that the accuracy dropped slightly when

these bands were removed. Hence all the

multispectral bands (bands 1-7, 9) of LANDSAT8

OLI data were used in the classification except the

thermal bands (TIRS 1&2).

Neural Net classifier in ENVI 5.2 was used to

apply a multi-layered feed-forward neural network

classification. An eleven layer multi-layered Neural

Network has been used for this eight-class satellite

image classification problem. The input layer

consists of eight neurons representing the eight

bands of the multi-spectral data. The output layer

has eleven neurons, representing the eleven crop and

cover type classes. For this study, we used only a

single hidden layer perceptron network based

classifier, with eight neurons in the hidden layer.

ENVI implementation of Neural Net allows

choosing between a logistic or hyperbolic activation

function. A logistic activation function was selected

due to its superior performance over the hyperbolic

function. There are four important parameters that

need to be set; namely training threshold

contribution, training rate, training momentum and

RMS exit criteria. By a process of iteration to

optimize these parameters, default values set by

ENVI were found to give best results except for

training threshold contribution that gave best

performance when it was set to 0.65. These values

were employed for classification for WorldView-3

and LANDSAT8.

Support Vector Machine (SVM) is a supervised

classification method derived from statistical

learning theory that often yields good classification

results from complex and noisy data (Pal and Foody,

Investigating the Use of High Resolution Multi-spectral Satellite Imagery for Crop Mapping in Nigeria - Crop and Landuse Classiﬁcation

using WorldView-3 High Resolution Multispectral Imagery and LANDSAT8 Data

113

2012, Jia et al., 2014). It separates the classes with a

decision surface that maximizes the margin between

the classes. The surface is often called the optimal

hyperplane, and the data points closest to the

hyperplane are called support vectors. The support

vectors are the critical elements of the training set.

SVM can be adapted to become a nonlinear

classifier through the use of nonlinear kernels. The

radial basis function Kernel (RBF) is the default in

ENVI. This has been found to give the best results

by many authors (Hsu et al., 2010, Jia et al., 2013).

The RBF kernel non-linearly mapped samples into a

higher dimensional space so the RBF could handle

the case when the relationship between class types

and attributes was not linear. Second, the RBF

kernel had fewer numerical computational

difficulties. The penalty value C and kernel

parameter γ were the two parameters used for the

RBF kernels, set to default values of 100 and 0.125

respectively as a result of our iteration process to

optimize them.

5 RESULT AND DISCUSSION

Figure 2 presents the results of the three algorithms

in classifying the crops and land cover types in the

study area. Clearly, forest and degraded forest land

cover types dominate the entire area. Primary forest

occurred mainly in the northern and western part of

the land while degraded forest existed mainly in the

south eastern part of the land. Fallow/grassland

appeared to be third largest among the land cover

type and it spread mainly around the cropped area.

Generally all three supervised classification

techniques seem to agree in discriminating the land

cover types in the area by visual interpretation of

figure 2(a)-2(c). A closer view of the classification

result is presented in Figure 3 (a-k). Figure 3a-d

presents a natural colour image of WorldView-3

image taken on January 3, 2015 and the

classification results of all three algorithms of an

area planted with young cassava and very young

cassava. Clearly all three classifiers were able to

discriminate young cassava class better than they did

very young cassava. Very young cassava category

was mainly confused with the bare ground/dirt road

class expectedly since this class had little vegetation

cover. This result implies that WorldView-3

multispectral products can identify cassava when it

is above 3 months old.

Although the three classification techniques were

able to distinguish from other land cover types such

as degraded forest and forest, they could not clearly

discriminate them from fallow/grassland. This is

probably due to spectral similarities between

fallow/grassland and matured cassava category

which are mainly mixtures of perennial shrubs and

weeds. LANDSAT8 OLI classification results of the

same matured cassava area present a better

performance (Figure 3i-k). With LANDSAT8, the

coarser spatial resolution smoothed out the noisy

pixels and produced more realistic results from the

three classifiers for the matured cassava farm area.

This result is similar to what Yang et al, 2007

obtained while they aggregated Quick Bird imagery

of spatial resolution of 2.8m to 11.2, 19.6 and 30m

from 2.8m to 11.2 m and 19.6 pixel sizes improved

overall classification accuracy in crop identification

in South Texas. LANDSAT8 of 30 m pixel size

identified matured cassava farm more realistically

than WorldView-3 at 1.6m spatial resolution.

0 to 0 present the accuracy assessment confusion

matrix for WorldView-3 image classification by

Maximum likelihood classifier (MLC), Neural Net

Classifier (NNC) and Support vector Machine

classifier (SVM). Overall accuracy of the three

classifiers are higher than 90 % with NNC having

the highest accuracy of 92% followed by MLC. The

kappa coefficients are equally high, greater than 0.81

for all three classifiers, though NNC still had the

highest at 0.833. The high kappa coefficients

indicate that all three classifiers performed at over

80% better than if the pixels have been randomly

assigned.

The producer’s accuracy is a measure of

omission error and it indicates the probability that

pixels that belong to the ground truth class and the

classification technique has failed to classify them

into the proper class. This ranged from 44.6% to

99.8% for MLC, 7.3% to 99.6% for NNC and for

SVM 12.3% to 99.3%. The lowest producer’s

accuracy occurred for the very young cassava

category for all the three classifiers indicating the

most difficult class to identify is the very young

cassava, although MLC performed best in

classifying this category at about 45% accuracy. The

highest confusion with this class came from young

cassava category and followed by matured maize for

MLC while it was confused mostly with matured

maize and bare ground classes for NNC. SVM

confused this class with bare ground class mostly

and closely followed by matured maize.

This is not unexpected since this class has a lot

of bare ground in the class due to sparse vegetation

of cassava at this stage of less than 2 months old.

Confusion with matured maize is probably due to

the fact that some of the matured maize plots

GISTAM 2016 - 2nd International Conference on Geographical Information Systems Theory, Applications and Management

114

Figure 3: (a) WorldView-3 natural colour image taken January 3, 2015, (b) Maximum Likelihood, (c) Neural Net, (d)

Support vector machine classification of an area of young cassava and very young cassava. (e) WorldView-3 natural colour

image, (f) Maximum Likelihood, (g) Neural Net, (h) Support vector machine classification of a matured cassava farm area.

(i) Maximum likelihood, (j) Neural Net and (k) support vector machine of the same matured cassava farm area using

LANDSAT8, OLI captured on January 15 2015.

have been harvested at the time of image capture

exposing more bare ground in these plots. From

Table 2-4, it is also clear that the three classifiers

identified the Forest, Built up, Major River and

Tarred roads categories correctly at a producer’s

accuracy of higher than 97% implying that the

probability of using WorldView-3, 8- band

multispectral image to identify those classes are

above 95%. A similar look at the user’s accuracy,

which is a measure of commission error and

indicative of the probability that a category

classified on the map actually represents that on the

ground reveals that it ranged from 35 to 100% for

MLC, 64 to 100% for NNC and 44 to 100% for

SVM. Clearly four of the eleven land cover

categories (Forest, Major River, Tarred Road and

Built up) were the easiest to identify with both user’s

and producer’s accuracy higher than 95% for all the

three classifier algorithms. The low producer’s

accuracy and user accuracies for the three cassava

categories (very young cassava, young cassava and

matured cassava) suggests that cassava crop is the

most difficult to differentiate among the eleven

landcover categories. The very young cassava and

young cassava were mostly confused with bare

ground, matured maize and fallow/grassland by the

3 algorithms although MLC produced least

confusion. Young cassava as well was often

confused with bare ground and matured maize. This

is because these two cassava crop categories have

significant bare ground exposure. The matured

cassava category was confused mainly with

degraded forest and Fallow/grassland under MLC

classifier, while it was mainly confused with

Fallow/grassland for NNC and SVM classifiers.

Spectral similarity between matured cassava and

fallow/grassland is expected since both consist of

shrubby vegetation and grasses. Most of the matured

cassava plots were also mixtures of weeds and

cassava which is a shrubby crop. This is the normal

practice in West Africa where weeds in cassava

farms of over 10 months are no longer controlled

since the farmers know the weeds competition with

cassava at this stage is very negligible.

MLC correctly identified matured cassava at a

producer accuracy of 83% whereas only 63% of

those pixels called matured cassava on the map are

actually matured cassava on the ground. Similarly

NNC identified this class at producer accuracy of

70% with a user accuracy of 67%, while SVM

Investigating the Use of High Resolution Multi-spectral Satellite Imagery for Crop Mapping in Nigeria - Crop and Landuse Classiﬁcation

using WorldView-3 High Resolution Multispectral Imagery and LANDSAT8 Data

115

Table 2: Confusion matrixes for crop and landcover classification of the WorldView-3 using the Maximum Likelihood

classifier.

Class