Computers and Electronics in Agriculture 166 (2019) 105018

Contents lists available at ScienceDirect

Computers and Electronics in Agriculture

journal homepage: www.elsevier.com/locate/compag

Land suitability assessments for yield prediction of cassava using geospatial T

fuzzy expert systems and remote sensing
Riska Ayu Purnamasaria, Ryozo Noguchib, Tofael Ahamedb,
⁎

a
Graduate School of Life and Environmental Sciences, University of Tsukuba, Japan
b
Faculty of Life and Environmental Sciences, University of Tsukuba, Japan

ARTICLE INFO ABSTRACT

Keywords: Cassava has the potential to be a promising crop that can adapt to changing climatic conditions in Indonesia due
Land suitability to its low water requirement and drought tolerance. However, inappropriate land selection decisions limit
Cassava cassava yields and increase production-related costs to farmers. As a root crop, yield prediction using vegetation
Yield prediction indices and biophysical properties is essential to maximize the yield of cassava before harvesting. Therefore, the
Fuzzy expert systems
purpose of this research was to develop a yield prediction model based on suitable areas that assess with land
Remote sensing
suitability analysis (LSA). For LSA, the priority indicators were identified using a fuzzy expert system combined
with a multicriteria decision method including ecological categories. Furthermore, the yield prediction method
was developed using satellite remote sensing datasets. In this analysis, Sentinel-2 datasets were collected and
analyzed in SNAP® and ArcGIS® environments. The multisource database of ecological criteria for cassava
production was built using the fuzzy membership function. The results showed that 42.17% of the land area was
highly suitable for cassava production. Then, in the highly suitable area, the yield prediction model was de-
veloped using the vegetation indices based on Sentinel-2 datasets with 10 m resolution for the accuracy as-
sessment. The vegetation indices were used to predict cassava growth, biophysical condition, and phenology
over the growing seasons. The NDVI, SAVI, IRECI, LAI, and fAPAR were used to develop the model for predicting
cassava growth. The generated models were validated using regression analysis between observed and predicted
yield. As the vegetation indices, NDVI showed higher accuracy in the yield prediction model (R2 = 0.62)
compared to SAVI and IRECI. Meanwhile, LAI had a higher prediction accuracy (R2 = 0.70) than other bio-
physical properties, fAPAR. The combined model using NDVI, SAVI, IRECI, LAI, and fAPAR reported the highest
accuracy (R2 = 0.77). The ground truth data were used for the evaluation of satellite remote sensing data in the
comparison between the observed and predicted yields. This developed integrated model could be implemented
for the management of land allocation and yield assessment in cassava production to ensure regional food
security in Indonesia.

1. Background land suitability evaluation based on FAO classes (FAO, 1976). To ac-
commodate the complexity in assigning suitability classes, the fuzzy
The productivity of cassava is one of the most important issues re- expert system combined with MCDM has the potential to determine the
lated to national food policy in Indonesia. Inappropriate management land suitability classification for cassava production. In MCDM, fuzzy
decisions and unsuitable land selection could limit productivity. Land set membership can be used in to standardize the criteria (Aydi et al.,
suitability assessment is essential for current and future land use 2016; Feizizadeh and Blaschke, 2013; Romano et al., 2015). The fuzzy
planning. The land suitability assessment can be performed using the set theory allows for continuous factors to be modeled for a suitability
multicriteria decision method. Several approaches of the multicriteria assessment within a GIS analysis.
decision method (MCDM) using expert systems have been used to However, most of those studies either used the MCDM technique or
conduct land suitability assessments (Malczewski, 2006; Ferretti and fuzzy set alone, resulting in a poorly handled weight of each factor or an
Pomarico, 2013; Elsheikh et al., 2013). The MCDM becomes more ap- inappropriate calculation of the suitability index. Furthermore, the
propriate with geospatial references. In recent years, computing tech- MCDM with fuzzy set theory has the potential to reduce the subjectivity
nologies combined with GIS have enabled geospatial references using in the assessment of results. The integrated approach of GIS, fuzzy set,

⁎
Corresponding author.
E-mail address: [email protected] (T. Ahamed).

https://doi.org/10.1016/j.compag.2019.105018
Received 10 October 2018; Received in revised form 3 July 2019; Accepted 19 September 2019
Available online 26 September 2019
0168-1699/ © 2019 Elsevier B.V. All rights reserved.
R. Ayu Purnamasari, et al. Computers and Electronics in Agriculture 166 (2019) 105018

Indonesia

Banten Province

Fig. 1. Land suitability analysis in the Banten province of Indonesia.

and MCDM has great potential to increase the effectiveness and accu- that requires yield prediction based on the canopy and biomass devel-
racy of land suitability assessment for crop production (Aguilar-Rivera opment over the growing seasons. To the best of our knowledge, there is
et al., 2018). Moreover, the effectiveness and accuracy of land suit- no research reported on cassava yield prediction at the suitable loca-
ability assessment can give more advantages for the yield prediction tions for cassava production. The yield prediction also helps in regional
model. Because, a yield prediction model based on reference points food security policy and production inventories to understand the
such as highly suitable area may have higher accuracy than statistically availability of cassava. Thus, the objective of this research is to develop
predicted yields used in previous studies (Feizizadeh and Blaschke, a model from various vegetation indices and biophysical properties
2013; Romano et al., 2015). derived from satellite datasets to predict the yield of cassava within
Furthermore, considering cassava is a root crop, the prediction of highly suitable area using MCDM and fuzzy expert system.
yield using vegetation indices (VIs) and biophysical properties is im-
portant to maximize yield before harvesting. The most straightforward
approach to estimating crop yields is to establish empirical relation- 2. Methodology
ships between ground-based yield measures and VIs measured on a
single date or integrated over the growing season (Tucker, 1979). In 2.1. Study area
this regard, satellite remote sensing and GIS applications for monitoring
crops have the potential for timely assessments of changes in the Banten Province is located between 5°7′50″–7°1′1″S and
growth and development of crops on regional scales (Lobell et al., 2015; 105°1′11″–106°7′12″E having 9663 km2 of total area and is the most
Campos et al., 2018). The rapid development of spectral reflectance and western point of Java, approximately 90 km from Jakarta. Banten
remote sensing technologies has been used in various models to esti- Province is strategically positioned as the area of land connecting Java
mate crop yield using remote sensing applications (Azzari et al., 2017; and Sumatra Islands. Since the productivity of cassava is decreasing in
Lobell et al., 2015; Sakamoto et al., 2013; Zabihi et al., 2015). For Indonesia, the productivity of cassava in Banten province should be
example, healthy crops and stress conditions are distinguished by the increased to support national food security (Fig. 1). Cassava production
absorption of red radiation and the reflectance of NIR radiation in the Banten province has the strong historical aspect. In some regions
(Tucker, 1979). Moreover, in the further research, the combination of of Banten, cassava is an important alternative source of food, especially
the reflectance’s is more complex, not only red and near-infrared re- for traditional cuisine that is prepared for traditional events. Regarding
flectance, but also green, red-edge and others. This combination also this condition, the research for cassava is very important for the Banten
defined as vegetation indices (VIs) (Frampton et al., 2013). province. The research conducted initially for the Banten Province and
However, there is a lack of research on cassava, which is a root crop can be expended in other parts of Java Island.

2
R. Ayu Purnamasari, et al. Computers and Electronics in Agriculture 166 (2019) 105018

Fig. 2. Framework of the research.

2.2. Framework of land suitability analysis references and a literature review. In ArcGIS 10, there are seven kinds
of fuzzy functions. In this study, based on the ecological criteria in the
As depicted in Fig. 2, the model was built over several stages. First, Banten province, we selected four functions: Large, Small, Gaussian and
in the land suitability analysis, all the criteria of ecological potential of Linear functions. These functions produce continuous fuzzy classifica-
Banten province were standardized using fuzzy membership functions. tions of standardized criteria.
Second, the highly suitable area for cassava was obtained in the The Fuzzy Large transformation function is used when the larger
ArcGIS® environment using weighted overlay tools and influenced by input values are more likely to be a member of the set. The Fuzzy Small
weights from MCDM. Then, cassava yield data points within the highly transformation function is used when the smaller input values are more
suitable areas were collected. These points were correlated to the cor- likely to be a member of the set. The Gaussian function is useful if the
responding vegetation indices (VIs) and biophysical properties to gen- membership functions represent the uncertainty in measurement.
erate a regression model for the early prediction of cassava yield for the Whereas, the Fuzzy Linear transformation function applies a linear
given fields. The Sentinel-2 satellite data sets were obtained to analyze function between the user-specified minimum and maximum values
the VIs and biophysical properties. (ESRI, CA, USA).
In the Fuzzy Large, Small and Gaussian membership functions, the
2.3. Land suitability analysis (LSA) control point must include a midpoint (f2) and spread (f1) (Table 1). A
midpoint is a point that has a 0.5 value of membership in Large, Small,
2.3.1. Reclassification by fuzzy membership function and Gaussian functions and is determined by the user based on refer-
The fuzzy set theory allows for the concept of these continuous ences (ESRI, CA, USA). The spread is generally assigned a number be-
factors to be modeled in a suitability assessment within a GIS or spatial tween 1 and 10. A higher spread value causes the fuzzy membership
domain (Fig. 3). In a standard approach, membership within a set or curve to become steeper. Meanwhile, the fuzzy linear transformation
class, is clearly and crisply defined as either in the class or not in the function applies a linear function between the minimum and maximum
class (Bellman and Zadeh, 1970). In the present study, fuzzy member- values. Anything below the minimum will be assigned a 0 (not a
ship classification is used to accommodate the high uncertainty of member) and anything above the maximum a 1 (a member) (Figs. 4 and
scoring methods in assigning the suitability classes. In this study, fuzzy 5). The equations for the fuzzy Large (Eq. (1)), Small (Eq. (2)), Gaussian
membership functions were used in ArcGIS 10® (ESRI, CA, USA) for (Eq. (3)), and Linear functions (Eq. (4)) are,
standardization. Several fuzzy functions were determined based on

3
R. Ayu Purnamasari, et al. Computers and Electronics in Agriculture 166 (2019) 105018

1.2

Highly Suitable (S1)

0.8
Fuzzy Membership

Moderately Suitable (S2)

0.6

0.4
Marginally Suitable (S3)

0.2
Not Suitable (N)

0
0 20 40 60 80 100 120
Input Value
Fig. 3. Reclassification by fuzzy membership functions.

µ (x ) =
1 on slopes between 0% and 15%. These slope ranges were determined to

() be highly suitable membership values. Based on the case and the

f1
x
1+ f2 (1) criteria, we selected the small membership function that explained the
1 small slope values indicating high suitability. The control point had a
µ (x ) = midpoint value of 25 and a spread value of 5.
()
f1
x
1+ f2 (2)

µ (x ) = e ( f 1 (X f 2) ) 2
(3) 2.3.2.2. Rainfall. Banten is characterized by a tropical climate and has
different variability in some areas. This factor could affect the spatial
0 x a pattern of the cassava yield in the whole region. The average
x a
µ (x ) = f (x ) = b a
a <x<b temperature and rainfall levels are 27.4 °C and 1500–2000 mm/year,
1 x b (4) respectively (BPS, 2017). The smallest amount of water, cassava can
grow with is approximately 400 mm/year, but higher yields can be
obtained with a greater amount of water supplied to the area (FAO,
2.3.2. Geographical extent and climatic factor 2013). For this condition, the Gaussian fuzzy function was used. The
2.3.2.1. Slope. In Banten Province, most topography was classified as control point was 1500 mm/year with a spread of 10-4.
slopes between 0% and 45% in steepness. Cassava is easily cultivated

Table 1
Suitability class by fuzzy membership for cassava production.
Suitability classes Fuzzy membership function Equation

Criteria S1 S2 S3 N Mid-point Spread

Slope (%) 0–8% 8–15% 15–25% 25–45% 25 5 µ (x ) =

1
x 5
1+
25
-4
Rainfall (mm/year) 1500 1500–3000 < 1500 > 3000 1500 10 0.0001 (X 1500)2)
µ (x ) =e (
Elevation (m) 0–100 100–300 300–700 > 700 300 7 µ (x ) =
1
x 7
1+
300
NDVI 1–0.7 0.7–0.5 0.5–0.3 0.3–0 0.53 5 µ (x ) =
1
x 5
1+
0.53
-4
Distance from rivers (m) 1000 1000–1500 < 500 > 1000 1000 10 0.0001 (X 1000)2)
µ (x ) =e (
Distance from roads (m) < 1000 1000–2000 2000–3000 > 3000 2000 5 µ (x ) =
1
x 5
1+
2000

Criteria S1 S2 S3 N Minimum Maximum

LULC Agri-culture dry field Rice-field Open field Forest, Settle-ments, Waterbody 1 10 0 x 1
x 1
(x ) = f (x ) = 1 < x < 10
10 1
1 x 10
Soil Type Latosol Rego-sol Alu-vial Podzol 1 25 0 x 1
x 1
µ (x ) = f (x ) = 1 < x < 25
25 1
1 x 25

4
R. Ayu Purnamasari, et al. Computers and Electronics in Agriculture 166 (2019) 105018

1
1
0.8

Membership Values

Membership Values
0.8
0.6
0.6
0.4 0.4
0.2 0.2
0 0
0 10 20 30 40 0 1000 2000 3000 4000 5000
Slope % Rainfall (mm/year)
(a) Slope membership function (b) Rainfall membership function

1 1

0.8 0.8

Membership Values
Membership Values

0.6 0.6

0.4 0.4

0.2 0.2

0 0
0 200 400 600 800 0 0.2 0.4 0.6 0.8 1
Elevation (m) NDVI Value

(c) Elevation membership function (d) NDVI membership function

1 1

0.8 0.8
Membership Values

Membership Values

0.6 0.6

0.4 0.4

0.2 0.2

0 0
0 500 1000 1500 2000 0 1000 2000 3000
Distance from river (m) Distance from road (m)
(e) Distance from river membership function (f) Distance from road membership function

1 1
Membership Values

Membershp Values

0.8 0.8
0.6 0.6
0.4 0.4
0.2 0.2
0 0
0 2 4 6 8 10 0 5 10 15 20 25
LULC Values Soil Values
(g) LULC membership function (h) Soil membership function

Fig. 4. (a-h) Fuzzy membership function criteria.

5
R. Ayu Purnamasari, et al. Computers and Electronics in Agriculture 166 (2019) 105018

(a) Soil (b) Elevation

(c) Rainfall (d) LULC

(e) NDVI (f) Distance from road

(g) Distance from river (h) Slope

Fig. 5. (a-h) Reclassification of the criteria.

6
R. Ayu Purnamasari, et al. Computers and Electronics in Agriculture 166 (2019) 105018

Table 2 determined as a linear function of fuzzy membership. The minimum

Ground reference collection for cassava yield within highly suitable area. membership value referred to a waterbody surface, and the maximum
Field ID Village Latitude Longitude Yield (ton/ha) referred to a dry agricultural field surface. The LULC database was
divided into ten classifications.
31 Cililitan −6.51 105.98 12.00
33 Pasireurih −6.40 105.94 13.00
57 Bojongjuruh −6.53 105.99 23.00
2.3.2.8. Soil. The soil plays a significant role in crop growth and
58 Sumberwaras −67.56 105.99 17.30 cassava yield (Howeler, 1991). The major soil types found in Banten
61 Bejod −6.76 105.94 16.00 are alluvial, red regosol, red-yellow podzolic, and latosol soils.
63 Ciginggang −6.55 106.02 21.00 According to Wargiono and Sudaryanto (2000), latosol areas are
65 Kramat Jaya −6.62 106.07 9.25
optimal for cultivating cassava. Latosol soils have good physical
66 Cijaku −6.70 106.06 25.00
69 Lebak tipar −6.96 106.34 15.00 properties and are deep and tolerant to erosion. However, podzols
70 Peucang pari −6.72 106.13 15.00 include low levels of organic matter and tend to erode easily. Just as
71 Cirinten −6.67 106.19 16.00 LULC was classified, the soil was classified with the linear function,
76 Panyaungan −6.91 106.16 20.00 with Latosol as the maximum value and podzols as the minimum value.
79 Pabubulan −6.92 106.29 22.00
80 Lebak −6.96 106.34 6.00
82 Cimayang −6.60 106.17 5.00 2.3.3. Weighted linear combination (WLC)
83 Cimayang −6.59 106.17 22.00 The WLC method was used to calculate the suitability index (S) for
84 Bojong −6.58 106.20 10.26 the area, Wi is the relative importance weight of criterion, Ri is the
85 Cisimet −6.57 106.24 8.00
86 Cisimet Raya −6.56 106.31 17.00
standardized value of the area under criterion i and n is the total
91 Sukamarga −6.53 106.32 23.00 number of criteria. The suitability score for each of the land units was
92 Padasuka −6.31 106.20 3.33 calculated using the following equation:
93 Bojongleles −6.36 106.21 5.80
n
94 Bojong cae −6.31 106.25 16.00
98 Sangiang −6.38 106.41 15.00
S= Wi × Ri
i=1 (5)
101 Cileles −6.48 106.11 4.00
In the FAO's framework for land evaluation, the land index classi-
fication was designated as suitable (S) or not suitable (N). These suit-
2.3.2.3. Elevation. In Indonesia, most cassava-growing areas are ability classes can then be further subdivided as required. The suit-
located in the lowland humid and subhumid tropics (Heumann et al., ability classification aimed to show the suitability of each land unit for
2011). In some areas, cassava can be grown in hilly or mountainous cassava production. In this section, three classes (S1, S2, and S3) were
terrain. In Banten Province, elevation ranges from 0 to 700 m. Most of used to distinguish land that is highly suitable, moderately suitable and
the area is suitable for cassava production, although the optimal marginally suitable for cassava production.
elevation for cassava production is approximately 62.5–137.5 m. In For the verification of the F-MCDM model, we used the regression
this case, the small function was fitted with a control point midpoint of analysis using ground truth data collected from the cassava yield within
300 m and a spread of 7. all suitable areas in the Banten province. The correlation of the Land
Suitability Index (LSI) with actual yield was developed (Azzari et al.,
2.3.2.4. Vegetation index. The vegetation canopy is an important factor 2017; Braimoh et al., 2004; Sys et al., 1991).
that could prevent soil erosion during cassava production. The NDVI is
a vegetation index that is correlated with several important biophysical 2.4. Yield prediction
properties and is used to generate different crop indices (Elhag, 2014).
Higher vegetation canopy indicated more highly suitable areas for 2.4.1. Sentinel-2 imagery
cassava production. In this circumstance, the large function was used All 24-available satellite images from Sentinel-2A over Banten
with 0.53 as the midpoint and a spread of 5 for the control point. province taken between March 2017 and February 2018 were down-
loaded from the USGS earth explorer internet resource. The cloud
2.3.2.5. Distance from river. In Banten Province, there are four wide masking and buffering were performed using idepix toolbox of Sentinel
river regions (Cibaliung, Ciliman, Cidanau, and Cisadane) that include Application Platform (SNAP) 5.0. For calculating the vegetation indices,
over 102 watershed areas. The physical factors associated with water several bands at a 10-meter resolution were used. The bands included in
supply were used to determine suitability levels for cassava production. this analysis were Band 4 (red), Band 8 (NIR) and three new bands in
Cassava can be planted on areas that are located not too far or too close the red-edge region, Band 5, Band 6 and Band 7, which are centered at
to water resources. The Gaussian function was chosen when the 705, 740 and 783 nm, respectively.
midpoint referred to higher suitability. The control point was 1000 m
for its midpoint and 10−4 for its spread. 2.4.2. Vegetation indices
2.4.2.1. NDVI. The NDVI was proposed by Rouse et al. (1973) and it
2.3.2.6. Distance from road. In selecting areas suitable for cassava has become the most widely used indicator for studying vegetation
production, the distance from roads must be considered because such canopy, crop production, green biomass, chlorophyll content, and
distances affect transportation costs for supply processes. Shorter canopy water stress (Vrieling et al., 2011). The NDVI utilizes two
distances between fields and roads facilitate access to transportation essential wavelength bands: the red and near-infrared (NIR) bands. The
infrastructure and link farmers and farming activities to marketing combination of Band 4 and Band 8 were used, corresponding to red and
channels. This means that the small function was fitted for this NIR reflection, respectively. The relationship between NDVI and yield
criterion. The control point was set with a midpoint of 2000 m and a has been described in various experiments. The NDVI was calculated as
spread of 5. follows:
RNIR RRed
2.3.2.7. Land use and land cover. Land use and land cover (LULC) data NDVI =
RNIR + RRed (6)
files describe the vegetation, water, natural surfaces, and cultural
features of a land surface (Akıncı et al., 2013). Rice fields covered
most lands in the Banten province. The classifications of LULC was 2.4.2.2. SAVI. The calculation of vegetation indices is influenced by

7
R. Ayu Purnamasari, et al. Computers and Electronics in Agriculture 166 (2019) 105018

0.7

NDVI SAVI
0.6

0.5
Vegetation Indices
0.4

0.3

0.2

0.1

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29
Number of fields
(a) Variability of NDVI and SAVI in Banten province

30

25

20
Yield (Ton/Ha)

15

10

0
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29
Number of fields
(b) Variability of yield of cassava in Banten province
Fig. 6. Variability of cassava fields in the same growing season.

soil background conditions in some cases (Huete, 1988; Gilabert et al., where RRE1 (Red Edge 1) is Band 5 (705 nm) and RRE2 (Red Edge 2) is
2002). Huete (1988) used a soil adjustment factor, L, to reduce the soil Band 6 (740 nm).
background effect. The proposed constant adjustment factor (L = 0.5)
is referred to as the soil-adjusted vegetation index (SAVI):
2.4.3. Biophysical properties
R RRed 2.4.3.1. LAI. The biophysical monitoring of essential plant properties,
SAVI = (1 + L) x NIR
RNIR + RRed (7) such as chlorophyll, nitrogen, LAI, leaf water content and crop health, is
very important. The leaf area index (LAI) is an important biophysical
variable for agricultural land monitoring and modeling studies. Since it
2.4.2.3. IRECI. IRECI (Inverted Red Edge Chlorophyll Index) utilized plays a important role in ecological processes, LAI retrieval maps from
all red-edge bands provided by Sentinel-2. The combined bands have satellites have been used in many land observation studies. LAI retrieval
substantial effects on chlorophyll absorption and internal leaf from Sentinel-2 was calculated using an algorithm provided by the
scattering. The IRECI also has potential as the optimal red edge index SNAP® biophysical operations toolbox.
for evaluating the biomass status using Sentinel-2A imagery (Castillo
et al., 2017). 2.4.3.2. fAPAR. Some research found that total biomass production is
(RNIR RRed) closely related to the fraction of photosynthetically active radiation
IRECI = RRE 2 (fAPAR). fAPAR is absorbed in the canopy over the growing season.
RRE1 (8) Estimations of fAPAR are often derived from Vis. In the SNAP®

8
R. Ayu Purnamasari, et al. Computers and Electronics in Agriculture 166 (2019) 105018

0.7

0.6

Vegetation Indices 0.5

0.4
NDVI
SAVI
0.3
IRECI

0.2

0.1

0
06/03/2017 14/06/2017 24/07/2017 01/11/2017 26/12/2017 19/02/2018

(a) Vegetation Index values for cassava during growing season

2.5

2
Biophysical Properties

1.5

LAI
fAPAR
1

0.5

0
06/03/2017 14/06/2017 24/07/2017 01/11/2017 26/12/2017 19/02/2018

(b) Biophysical properties for cassava during growing season

Fig. 7. The optimum growing season based on VIs (a) and biophysical properties (b).

software, the calculation of fAPAR is also carried out using the

biophysical operations toolbox.

2.4.4. Ground reference data

The location points of the cassava fields were collected by GPS
Garmin eTrex (Garmin Inc. KA, USA) around Banten province in
2016–2017. Yield estimates were evaluated using the ground-truth
yield data based on the farm surveys at the cassava field locations. The
harvested time was recorded for each cassava field. The field location
and yield data collected on the highly suitable area listed in Table 2. In
the Banten province, the farmers have different times to start to plant
cassava. The variability of the fields within the various growing seasons
is shown in Figs. 6 and 7.

Fig. 8. Suitable area for cassava production.

9
R. Ayu Purnamasari, et al. Computers and Electronics in Agriculture 166 (2019) 105018

Table 3 biophysical properties had higher prediction accuracy (R2 = 0.70) than
Suitable areas for cassava production. fAPAR. The combined model using NDVI, SAVI, IRECI, LAI, and fAPAR
Suitability classes Fuzzy-MCDM reported the highest accuracy (R2 = 0.77). The combination model was
used to generate the yield prediction map. The ground truth data were
Percentage area (%) Area (ha) used to evaluate the satellite remote sensing data between the observed
and predicted yields. Finally, the yield map was generated to show the
Highly suitable 42.17 407,488
Moderately suitable 43.10 416,475 variability of cassava yield in the various suitable area in Banten pro-
Marginally suitable 6.25 60,393 vince (Fig. 12). The higher yield prediction was observed in the
Not suitable 8.47 81,845 southern part of the Banten province. This result was consistent re-
garding the highly suitable area that also located in the same region
(Fig. 8).
3. Results The prediction formula was applied for estimating the yield before
harvesting around December 2017. For comparison, the prediction was
3.1. Land suitability analysis performed in the cassava fields within highly suitable and moderately
suitable areas. In a highly suitable area, the average yield for cassava
The results showed that 42.17% of land area was highly suitable for production was 12.21 ton/ha. This prediction was much higher than
the F-MCDM model (Fig. 8 and Table 3). In this regard, F-MCDM land the prediction of cassava yield in the moderately suitable area (8.56
suitability maps reflected interaction between the fuzzy membership ton/ha) (Tables 5 and 6).
function values and their weights. To evaluate the accuracy of the
methods, the correlation between land suitability index (LSI) and actual
yield of cassava in various suitable area was obtained (Fig. 9). This 4. Discussion
regression shows that 55% of cassava fields in the study area were lo-
cated in suitable lands for cassava production. In the land suitability analysis, F-MCDM could be good approach
when the criteria were more complex and required a continuous value.
The F-MCDM method was validated by the fact that approximately 55%
3.2. Yield prediction using vegetation indices and biophysical properties of cassava fields were located in a proper suitable location. In the other
way we can assume the higher yield that located in higher suitability
The yield data from the field were compared with data retrieved area and lower yield that located in less suitable area is around 55%.
from Sentinel-2. The analysis was conducted to find the relationship The suitable areas were mostly located in the southern part of the
between yield, VIs, and biophysical values at each growing stage. This Banten Province. Therefore, farmers in the Banten province required
study investigated the use of VIs and biophysical properties in a more highly suitable land to cultivate cassava. In further assessments, the
detailed field area located in a highly suitable location based on a highly suitable areas were determined to predict the yield of cassava
ground-truth survey. The results of the study show that the VIs and production. Moreover, to help the farmer obtain more yield in the op-
biophysical properties during the 24 July 2017 growing season had timum window for harvesting, the prediction of cassava yield was
slightly stronger correlations than the other (Figs. 10, 11, and Table 4). needed. The best time to predict the cassava yield was observed be-
The yield prediction model was developed using the vegetation tween yield and vegetation indices in the green-up season around July
index retrieved from the Sentinel-2 datasets (10 m resolution). The 2017. This result indicated that cassava yield can be predicted in July,
vegetation indices were used to predict cassava growth, biophysical five months before schedule harvest in December or February (Fig. 11).
condition, and phenology over the growing seasons. The NDVI, SAVI, The vegetation phenology analysis and biophysical properties of cas-
IRECI, LAI, and fAPAR were used to develop the predictive model for sava plants has the potential for estimating yield prediction as root
cassava growth. The generated models were validated using regression crops with good accuracy in highly suitable areas. In general, all in-
analysis to estimate the variations among the observed and predicted dicators explain the vegetation canopy coverage. A good prediction
yields. As vegetation indices NDVI showed higher accuracy in the yield model was obtained from the combination of vegetation indices (NDVI,
prediction model (R2 = 0.62) compared to SAVI and IRECI. The LAI as SAVI, and IRECI) and biophysical properties (LAI, fAPAR). NDVI, LAI,

30

25

y = 0.1348x + 3.5722
Observed Yield (ton/ha)

20 R² = 0.55

15

10

0
0 20 40 60 80 100 120
Land Suitability Index
Fig. 9. Verification of the F-MCDM land suitability model.

10
R. Ayu Purnamasari, et al. Computers and Electronics in Agriculture 166 (2019) 105018

Fig. 10. Relationship between VIs, biophysical properties and the actual yield of cassava.

0.6

0.5

0.4 NDVI
SAVI
0.3 IRECI
R2

LAI
0.2 fAPAR

0.1

0
06/03/2017 14/06/2017 24/07/2017 01/11/2017 26/12/2017 19/02/2018
Fig. 11. Growing season for cassava in Indonesia.

11
R. Ayu Purnamasari, et al. Computers and Electronics in Agriculture 166 (2019) 105018

Table 4 5. Conclusions
The yield prediction models.
Vegetation indices R2 Yield expression Inappropriate decisions relating to land selection limit the pro-
ductivity of cassava cultivation and increase the associated costs to
NDVI 0.62 Yield = 39.234*NDVI − 3.6631 farmers. This research was conducted to develop a land suitability
SAVI 0.48 Yield = 49.895*SAVI + 0.4108
model on the provincial scale to find the best suitable areas for cassava
IRECI 0.37 Yield = 23.831*IRECI + 6.1878
production. Fuzzy set theory is advantageous for the standardization of
Biophysical R2 Formula criteria using fuzzy membership functions. In ArcGIS10®, there were 7
properties
LAI 0.70 Yield = 9.4269*LAI + 0.6379
fuzzy functions classified into Large, Small, Gaussian and Linear func-
fAPAR 0.27 Yield = 21.636*fAPAR + 2.19 tions. In other words, this research tried to demonstrate the sufficiency
All Combination 0.77 Yield = 33*NDVI – of fuzzy functions for cassava production area selection, which was
54*SAVI + 8*IRECI + 2.2*LAI + 4.8*fAPAR based on the midpoint and spread defined control points. After this
stage, the final suitability maps were determined using the weighted
overlay method. A suitable land map based on fuzzy-based MCDM was
generated. In further assessments, the highly suitable areas were de-
termined for predicting yield of cassava production.
The spectral bands of Sentinel-2 satellite imagery, vegetation in-
dices, and biophysical properties were used as input parameters to
generate cassava yield prediction models. The study was carried out in
cassava fields that were located in highly suitable locations for cassava
production in the Banten province in Indonesia. The regression models
were generated using the vegetative indices and biophysical properties.
The yield prediction model was developed using the vegetation indices
and biophysical properties from the Sentinel-2 datasets with 10 m re-
solution. The vegetation indices (NDVI, SAVI, IRECI) and biophysical
properties (LAI and fAPAR) were used to predict cassava growth, bio-
physical condition, and phenology over the growing seasons. The
ground truth data were used to evaluate of satellite remote sensing data
between the estimated yield within highly suitable and moderately
suitable areas. From this result, the estimation of cassava yield was
better when the field was located in a highly suitable area. The devel-
Fig. 12. Cassava yield prediction map (ton/ha). oped decision support system was integrated with the expert system,
GIS, and Sentinel-2 satellite datasets to evaluate land suitability and
predict cassava yield from canopy biomass and soil adjusted indices.
fAPAR related with chlorophyll content. Some related indices like SAVI
The developed model can be used at the regional and country levels in
and IRECI can be used to extract information on protein content and
land suitability assessment and yield estimation of cassava to maximize
crop yield prediction before the harvest time.
production. In further research, machine learning will be added to va-
lidate the yield prediction model. The developed integrated decision
support system model can be used to make recommendations to policy

Table 5
Yield prediction for cassava in highly suitable area.
Field ID Village Latitude Longitude Observed yield (ton/ha) NDVI SAVI LAI fAPAR IRECI Predicted yield (ton/ha)

31 Cililitan −6.51 105.98 12.00 0.58 0.37 2.08 0.65 0.49 10.72
33 Pasireurih −6.40 105.94 13.00 0.62 0.38 2.77 0.62 0.39 12.51
57 Bojongjuruh −6.53 105.99 23.00 0.60 0.41 1.81 0.61 0.51 8.90
58 Sumberwaras −68.00 105.99 17.30 0.78 0.54 3.76 0.83 1.12 18.18
61 Bejod −6.76 105.94 16.00 0.80 0.54 3.93 0.83 1.00 18.18
63 Ciginggang −6.55 106.02 21.00 0.76 0.54 3.97 0.84 1.08 17.36
65 Kramat Jaya −6.62 106.07 9.25 0.57 0.32 1.96 0.62 0.40 12.05
66 Cijaku −6.70 106.06 25.00 0.34 0.23 1.11 0.52 0.26 5.87
69 Lebak tipar −6.96 106.34 15.00 0.45 0.14 1.18 0.35 0.10 12.01
70 Peucang pari −6.72 106.13 15.00 0.64 0.42 1.94 0.64 0.59 10.60
71 Cirinten −6.67 106.19 16.00 0.37 0.21 1.08 0.44 0.22 6.71
76 Panyaungan −6.91 106.16 20.00 0.73 0.47 1.89 0.71 0.45 9.85
79 Pabubulan −6.92 106.29 22.00 0.75 0.49 2.14 0.71 0.86 13.42
80 Lebak −6.96 106.34 6.00 0.40 0.26 0.85 0.39 0.24 5.13
82 Cimayang −6.60 106.17 5.00 0.31 0.19 0.63 0.28 0.16 3.95
83 Cimayang −6.59 106.17 22.00 0.68 0.44 2.19 0.67 0.65 11.88
84 Bojong −6.58 106.20 10.26 0.73 0.47 2.64 0.75 0.94 15.40
85 Cisimet −6.57 106.24 8.00 0.70 0.42 2.44 0.70 0.66 14.31
86 Cisimet Raya −6.56 106.31 17.00 0.71 0.40 0.91 0.60 0.56 10.84
91 Sukamarga −6.53 106.32 23.00 0.79 0.46 2.75 0.72 0.79 17.09
92 Padasuka −6.31 106.20 3.33 0.82 0.57 3.70 0.83 1.24 18.34
93 Bojongleles −6.36 106.21 5.80 0.76 0.48 2.57 0.72 0.77 14.58
94 Bojong cae −6.31 106.25 16.00 0.61 0.37 1.53 0.54 0.47 9.69
98 Sangiang −6.38 106.41 15.00 0.66 0.41 2.00 0.64 0.53 11.71
101 Cileles −6.48 106.11 4.00 0.77 0.47 2.74 0.74 0.81 15.86
Average 12.21

12
R. Ayu Purnamasari, et al. Computers and Electronics in Agriculture 166 (2019) 105018

Table 6
Yield prediction for cassava in moderately suitable area.
Field ID Village Latitude Longitude Observed yield (ton/ha) NDVI SAVI IRECI LAI fAPAR Predicted yield (ton/ha)

37 Menes −6.39 105.93 13.00 0.60 0.23 0.25 0.88 0.36 13.02
40 Citalahab −6.38 106.11 12.00 0.35 0.33 0.40 1.08 0.61 2.53
87 Ciminyak −6.54 106.31 21.00 0.62 0.46 0.79 2.75 0.72 11.80
88 Sobang −6.62 106.30 10.62 0.48 0.57 1.24 3.70 0.83 7.03
100 Cikulur −6.40 106.18 12.00 0.52 0.39 0.62 1.94 0.66 8.42
Average 8.56

makers and planners in Indonesia to increase the practices of producing Elsheikh, R., Shariff, A.R.B.M., Amiri, F., Ahmad, N.B., Balasundram, S.K., Soom, M.A.M.,
cassava in highly suitable areas. Furthermore, the developed model can 2013. Agriculture Land Suitability Evaluator (ALSE): A decision and planning support
tool for tropical and subtropical crops. Comput. Electron. Agric. 93, 98–110. https://
be employed for yield prediction to develop an inventory planning doi.org/10.1016/j.compag.2013.02.003.
system to ensure regional food security. FAO, 1976. A Framework for Land Evaluation, first ed. FAO, Rome, Italy.
FAO, 2013. Save and Grow: Cassava A Guide to Sustainable Production Intensification.
FAO, Rome, Italy.
Acknowledgments Frampton, W.J., Dash, J., Watmough, G., Milton, E.J., 2013. Evaluating the capabilities of
Sentinel-2 for quantitative estimation of biophysical variables in vegetation. ISPRS J.
We would like to thank the University of Tsukuba to support this Photogramm. Remote Sens. 82, 83–92. https://doi.org/10.1016/j.isprsjprs.2013.04.
007.
research to develop the multi-criteria modeling for land suitability Feizizadeh, B., Blaschke, T., 2013. GIS-multicriteria decision analysis for landslide sus-
analysis for cassava production in Indonesia. We also express our sin- ceptibility mapping: comparing three methods for the Urmia lake basin, Iran. Nat.
cere thanks to the, Indonesian Geospatial Agency, the United States Hazards 65 (3), 2105–2128. https://doi.org/10.1007/s11069-012-0463-3.
Ferretti, V., Pomarico, S., 2013. An integrated approach for studying the land suitability
Geological Survey (USGS) and European Space Agency (ESA) for geo-
for ecological corridors through spatial multicriteria evaluations. Environ. Dev.
graphical and satellite data information. We sincerely thank the Sustain. 15 (3), 859–885. https://doi.org/10.1007/s10668-012-9400-6.
Indonesia Endowment Fund for Education (LPDP) for providing scho- Gilabert, M.A., González-Piqueras, J., Garcıa-Haro, F.J., Meliá, J., 2002. A generalized
larship to continue this research in Japan. We also expressed our gra- soil-adjusted vegetation index. Remote Sens. Environ. 82 (2–3), 303–310.
Huete, A.R, 1988. A soil-adjusted vegetation index (SAVI). Remote Sens. Environ. 25 (3),
titude’s to Indonesian experts and field surveyors to participate in this 295–309. https://doi.org/10.1016/0034-4257(88)90106-X.
research. Heumann, B.W., Walsh, S.J., McDaniel, P.M., 2011. Assessing the application of a geo-
graphic presence-only model for land suitability mapping. Ecol. Inf. 6 (5), 257–269.
https://doi.org/10.1016/j.ecoinf.2011.04.004.
References Howeler, R.H., 1991. Long-term effect of cassava cultivation on soil productivity. Field
Crops Res. 26 (1), 1–18. https://doi.org/10.1016/0378-4290(91)90053-X.
Aguilar-Rivera, N., Algara-Siller, M., Olvera-Vargas, L.A., Michel-Cuello, C., 2018. Land Lobell, D.B., Thau, D., Seifert, C., Engle, E., Little, B., 2015. A scalable satellite-based crop
management in Mexican sugarcane crop fields. Land Use Policy 78, 763–780. https:// yield mapper. Remote Sens. Environ. 164, 324–333. https://doi.org/10.1016/j.rse.
doi.org/10.1016/j.landusepol.2018.07.034. 2015.04.021.
Akıncı, H., Özalp, A.Y., Turgut, B., 2013. Agricultural land use suitability analysis using Malczewski, Jacek, 2006. GIS-based multicriteria decision analysis: a survey of the lit-
GIS and AHP technique. Comput. Electron. Agric. 97, 71–82. https://doi.org/10. erature. Int. J. Geographical Inf. Sci. 20 (7), 703–726. https://doi.org/10.1080/
1016/j.compag.2013.07.006. 13658810600661508.
Aydi, A., Abichou, T., Nasr, I.H., Louati, M., Zairi, M., 2016. Assessment of land suitability Romano, G., Dal Sasso, P., Liuzzi, G.T., Gentile, F., 2015. Multi-criteria decision analysis
for olive mill wastewater disposal site selection by integrating fuzzy logic, AHP, and for land suitability mapping in a rural area of Southern Italy. Land Use Policy 48,
WLC in a GIS. Environ. Monit. Assess. 188 (1), 59. https://doi.org/10.1007/s10661- 131–143. https://doi.org/10.1016/j.landusepol.2015.05.013.
015-5076-3. Rouse, J.W., Haas, R.H., Schell, J.A., Deering, D.W., 1973. Monitoring vegetation systems
Azzari, G., Jain, M., Lobell, D.B., 2017. Towards fine resolution global maps of crop in the Great Plains with ERTS, Third ERTS Symposium, NASA SP-351 I, 309–317.
yields: Testing multiple methods and satellites in three countries. Remote Sens. Sakamoto, T., Gitelson, A.A., Arkebauer, T.J., 2013. MODIS-based corn grain yield esti-
Environ. 202, 129–141. https://doi.org/10.1016/j.rse.2017.04.014. mation model incorporating crop phenology information. Remote Sens. Environ. 131,
Bellman, R.E., Zadeh, L.A., 1970. Decision-making in a fuzzy environment. Manage. Sci. 215–231.
17 (4), B-141. https://doi.org/10.1287/mnsc.17.4.B141. Sys, C., Van Ranst, E., Debaveye, I.J., 1991. Land evaluation. Part I: Principles In Land
BPS, 2017. Regional Statistics Banten Province 2017. Badan Pusat Statistik (The Central Evaluation and Crop Production Calculations. General Administration for
Statistics Agency of Indonesia). BPS Banten Province. Development Cooperation, Agricultural publication-No. 7, Brussels-Belgium, 274.
Braimoh, A.K., Vlek, P.L., Stein, A., 2004. Land evaluation for maize based on fuzzy set Tucker, C.J., 1979. Red and photographic infrared linear combinations for monitoring
and interpolation. Environ. Manage. 33 (2), 226–238. https://doi.org/10.1007/ vegetation. Remote Sens. Environ. 8 (2), 127–150.
s00267-003-0171-6. Vrieling, A., de Beurs, K.M., Brown, M.E., 2011. Variability of African farming systems
Campos, I., González-Gómez, L., Villodre, J., Calera, M., Campoy, J., Jiménez, N., Calera, from phenological analysis of NDVI time series. Clim. Change 109 (3–4), 455–477.
A., 2018. Mapping within-field variability in wheat yield and biomass using remote https://doi.org/10.1007/s10584-011-0049-1.
sensing vegetation indices. Precis. Agric. 1–23. https://doi.org/10.1007/s11119-018- Wargiono, J., Sudaryanto, B., 2000. Cassava leaves and forage crops for ruminant feed in
9596-z. the establishment of sustainable cassava farming system in Indonesia. In National
Castillo, J.A.A., Apan, A.A., Maraseni, T.N., Salmo III, S.G., 2017. Estimation and map- Workshop-Seminar on Sustainable Livestock Production on Local Feed Resources, pp.
ping of above-ground biomass of mangrove forests and their replacement land uses in 496–503. http://ciat.library.ciat.cgiar.org/Articulos_Ciat/proceedings_workshop_02/
the Philippines using Sentinel imagery. ISPRS J. Photogramm. Remote Sens. 134, 496.pdf.
70–85. https://doi.org/10.1016/j.isprsjprs.2017.10.016. Zabihi, H., Ahmad, A., Vogeler, I., Said, M.N., Golmohammadi, M., Golein, B., Nilashi, M.,
Elhag, M., 2014. Sensitivity analysis assessment of remotely based vegetation indices to 2015. Land suitability procedure for sustainable citrus planning using the application
improve water resources management. Environ. Dev. Sustain. 16 (6), 1209–1222. of the analytical network process approach and GIS. Comput. Electron. Agric. 117,
https://doi.org/10.1007/s10668-014-9522-0. 114–126. https://doi.org/10.1016/j.compag.2015.07.014.

13