Sentinel 2 High-Resolution Land Cover Mapping in Sub-Saharan

Africa with Google Earth Engine

Elena Belcore

and Marco Piras

DIATI, Politecnico di Torino, Corso Duca degli Abruzzi 24, 10129 Torino, Italy

Keywords: Land Cover, Machine Learning, Sentinel-2, Google Earth Engine, Sub-Saharan, Natural Hazard, Climate

Change.

Abstract: This work aims to develop an efficient methodology for high-resolution spatial and thematic land cover maps

of sub-Saharan areas based on Sentinel-2 data. LC mapping in these areas is complicated due to their land

morphology, climatic conditions and homogeneity of surface spectral responses. Two pixel-based supervised

classification approaches are compared in Google Earth Engine. The aggregated method classifies each image

and then aggregates the results on frequency bases at pixel level. The stacked method classifies all the images

together in a single stacked database. Additionally, the influence of linear atmospheric correction models on

the overall accuracy (OA) is assessed, and the best-performing approach is compared to existing Land Cover

(LC) maps of the area. 16 Sentinel-2 images (level 1C) from 2017 and 2019 were atmospheric and

topographically corrected and classified into nine classes. The results show similar performances for the

analysed approaches, with a slightly high OA for the aggregated classification (0.97). The atmospheric

correction has little impact on the results.

1 INTRODUCTION

Sub-Saharan Africa is exceptionally vulnerable to

Climate Change-induced phenomena, such as floods,

erosion, and droughts, which have dramatically

increased in the past years. The Dosso region in

southwest Niger is no exception (Bigi et al., 2018;

Oguntunde et al., 2018; Teodoro and Duarte, 2022).

In 2021, 200,000 people were affected by floods in

Niger. One of the most recent disastrous events

occurred in October 2022, when flooding caused by

heavy rains claimed nearly 200 lives and affected

more than a quarter of a million people.

In such scenarios, to plan against natural disasters,

continuous and detailed monitoring of the area is not

negligible (Tiepolo et al., 2018), and updated

information regarding the Land Cover (LC) is crucial

to land management, as these maps provide users with

information related to terrestrial ecosystems and

livelihoods (Li et al., 2020). Classification of Sub-

Saharan areas is one of the most challenging due to

the landscape complexity and the low spectral

variability within the covers. Moreover, the sand dust

https://orcid.org/0000-0002-3592-9384

https://orcid.org/0000-0001-8000-2388

particulates in the atmosphere may alter the spectral

response of the Earth surface and further exacerbate

the difficulty of the classification. One of the most

significant problems in the optical remote sensing of

Sub-Saharan regions is that reflectance from soil and

rock during the dry season is often much greater than

that of the sparse vegetation making it difficult to

separate the vegetation. Some specific problems

involved with remote sensing of arid vegetation

include multiple scattering of light (nonlinear mixing)

between vegetation and soil (Huete, 1988). Moreover,

the local architecture, such as Nigerienne

architecture, consists of tiny houses with flat roofs

built using local clays and not plastered, which results

in hardly spectral separable build-up areas and bare

soils, even from very high resolution (VHR) imagery

(Belcore et al., 2022), Figure 1.

Similarly, in some villages and suburbs, the roads

are unpaved. As a consequence, the spectral response

of buildings is the same as roads and bare soil. The

strong seasonality adds further complexity to the

classification as the frequent cloud cover during the

rainy season and sand presence in the air may alter the

spectral values of sensed data.

Belcore, E. and Piras, M.

Sentinel 2 High-Resolution Land Cover Mapping in Sub-Saharan Africa with Google Earth Engine.

DOI: 10.5220/0011746500003473

In Proceedings of the 9th International Conference on Geographical Information Systems Theory, Applications and Management (GISTAM 2023), pages 27-36

ISBN: 978-989-758-649-1; ISSN: 2184-500X

 2023 by SCITEPRESS – Science and Technology Publications, Lda. Under CC license (CC BY-NC-ND 4.0)

These reasons contribute to the LC data scarcity

in Sub-Saharan Africa. Indeed, to date, few very high-

resolution LC maps of sub-Saharan areas exist.

Examples are the Africa LC by ESA and FROM-

GLC10 (Li et al., 2020). Although these LC maps are

incredibly complex to realise, most do not provide

good thematic detail. Because of their nature, the

training datasets are hardly updatable due to the large

amount of time and manual work this task requires.

Figure 1: Aerial view of local architecture acquired by a

drone system in 2021.

Niger's southern territories are the study area of

the ANADIA 2.0 project (Anadia 2.0, 2023). It aims

to create an Early Warning System to face climate

change effects in Sirba River Basin in the Tillabery

region, enhance local technicians' knowledge

regarding floods forecasting, and create an adaptation

strategy planform two villages along the Sirba River.

In this area, the need for climatic planning and the

development of adaptation strategies to climate

change at the local level is not negligible (Tiepolo et

al., 2018). Despite this undeniable need for climate

planning, there is no appropriate risk mapping of the

area; indeed, subnational risk mapping lacks detail

(Tiepolo et al., 2018). The data gathered and the

information provided by this work are directly

involved in ANADIA 2.0 by feeding the adaptation

strategy plan and investigating the cause-effect

relation of floods.

Aiming to map and monitor the areas potentially

flooded by the Sirba river, a workflow to produce an

updated, high-detailed LC map of the site is proposed

and validated. In this application, the LC map of

southwest Niger has been realised using nine classes

over 16 features. The entire process was completed in

the Google Earth Engine (GEE) platform, and two

multi-temporal approaches for classification were

tested and compared.

2 MATERIALS AND METHODS

2.1 Study Area

Sirba River is a tributary of the Niger River and

crosses Burkina Faso and Niger countries (Figure 2).

Its basin is prone to floods, and villages along the

river are vulnerable to life and economic losses

(Massazza et al., 2019, 2018; Tamagnone et al.,

2019).

Figure 2: Footprint on the classification (red) and detail of

the Nigerienne branch of the Sirba River.

2.2 Identification of Classes

Nine classes describe the classification system, as

table 1 illustrates.

2.3 Satellite Imagery Filtering and

Pre-Processing

The images acquired by Sentinel-2 (both Sentinel-2A

and Sentinel-2B satellites) were filtered by location

and date. The study area includes the segment of

Sirba River that lies in Niger country for about 100

km. The period covers all the acquisitions between

2017 and 2019. An additional filtering parameter

regards the cloud cover percentage, which must be

less than 10% over a single scene. Only images

sensed during the rainy season (from August to

October) were selected to maximise the classes'

spectral variability, especially to better distinguish

between the vegetation classes and the bare soils and

to identify water. Sentinel-2 level 1C dataset was used

for this classification since only one image from the

corrected dataset of Sentinel-2 level 2A satisfied the

filter mentioned above parameters. The selected

images of Sentinel 2A-1C level are 16 (Table 2). The

GISTAM 2023 - 9th International Conference on Geographical Information Systems Theory, Applications and Management

dataset was atmospherically corrected by applying

Dark Object Subtraction (DOS) (Chavez, 1988),

which is a linear atmospheric correction model that

performs similarly to radiative transfer models on

homogeneous surfaces (Lantzanakis et al., 2017).

Table 1: Classes of the classification in South Niger.

N Class Description Picture

1 Water Internal waters

2 Plateaux

Elevated areas over the

dry savannah. They

influence the water

catchment, the erosion

process and present

peculiar plant species.

Riparian

vegetation

The thickly vegetated

area along the rivers. It

is usually composed of

trees and bushes.

4 Urban areas

Villages and main

roads.

Red bare

soils

Red soils rich in ferric

oxides that

characterised the

savannah soil

landscape.

Sandy bare

soils

Sand natural deposits.

Vegetation

of the

plateaux

Vegetation on the

plateaux. It grows

along the drainage

canals. It is mainly

composed of

herbaceous species.

Irrigated

agricultural

lands

Areas interested by

intense agricultural

activity that require

tillage, and irrigated

generally through

channel systems.

Non-

irrigated

agricultural

lands and

pastures

Areas interested by

moderate agricultural

activities that require

tillage or pastures.

Table 2: List of Sentinel-2 images used in the classification.

Yea

No. Sentinel Ima

e Identification Code

2017

0 20170815T102021

20170815T102513

T31PCR

1 20170924T102021_20170924T102649_T31PCR

2 20170926T101009

20170926T102049

T31PCR

2018

3 20180815T102019

20180815T102918

T31PCR

4 20180820T102021_20180820T103538_T31PCR

5 20180911T101019

20180911T101438

T31PCR

6 20180911T101019

20180911T102702

T31PCR

7 20180916T101021_20180916T101512_T31PCR

8 20180921T101019

20180921T101647

T31PCR

9 20180924T102019

20180924T102602

T31PCR

10 20180929T102021_20180929T103112_T31PCR

2019

11 20190812T101031

20190812T102016

T31PCR

12 20190911T101021

20190911T102116

T31PCR

13 20190921T101031

20190921T102426

T31PCR

14 20190926T101029

20190926T102551

T31PCR

15 20190929T102029

20190929T102700

T31PCR

Knowing that DOS can affect the classification's

results differently depending on the geographical area

(and land cover), the classification model was trained

and applied over the DOS-corrected and non-

corrected datasets. Then the Overall Accuracy (OA)

of the classifications were compared to check the

influence of DOS on the final result.

The topographical correction of the images was

initially applied to reduce the effects of elevation over

the plateaux areas (Dorren et al., 2003). The

topographic correction allows the variation in the

reflectance derived from the terrain's inclination and

the sun elevation (Poortinga et al., 2019; Shepherd

and Dymond, 2010). Nevertheless, the correction

introduced noise in the dataset because the plateaux

slopes do not interfere with the soil's spectral

response. The Digital Terrain Model (DTM) applied,

USGD 30m DTM, is not resolute and precise enough.

Thus, the dataset was not topographically corrected.

(a)

(b)

Figure 3: (a) RGB mosaic on sentinel-2, level 1C data, DOS

applied; (b) RGB mosaic on sentinel 1C data, DOS applied,

and topographically corrected. The correction excessively

alters the data over the plateaux and the plane areas.

Sentinel 2 High-Resolution Land Cover Mapping in Sub-Saharan Africa with Google Earth Engine

2.4 Features Extraction and Selection

The feature extraction consisted of the computation

of 6 spectral features, 4 histogram-based features, 18

textural (computed on Sentinel band 8A), 2 elevation-

derived features, and 1 edge-detector feature added to

the 12 spectral bands. Specifically, the Gray Level

Co-occurrence Matrix (GLCM) texture metrics were

calculated over a 9x9 neighbourhood (Haralick et al.,

1973), while the histogram-based features were on a

3x3 filter (Conners et al., 1984; GEE,2022) to help in

classes discrimination (Drzewiecki et al., 2013;

Kukawska et al., 2017). Table 3 lists the extracted

features.

Table 3: Derivative features calculated for Sentinel-2.

Feature Formula/note

Chlorophyll

IndexRedEdge, CRE

(B9/B5)−1

Enhanced Vegetation

Index, EVI

2.5*((B9−B5)/((B9+6*B5−7.5*B1)+1))

HUE

Arctan((2*V5−B3−B1)/30.5)*(B3−B1))

Soil Composition Index,

SCI

(B11−B8)/(B11+B8)

Wetness Index, WET

(0.1509*B2)+(0.1973*B3)+(0.3279*B4)+

(0.03406*B8)-(0.7112*B11)-(0.4572*B12)

Triangular Vegetation

Index, TVI

0.5*(120*(B8-B3))-(200*(B4-B3))

Sob Sobel edge extractor

Var Variance

Mean Mean

Skew Skewness

Kurt Kurtosis

Entr Entropy

Asm

Angular Second Moment; measures

the number of repeated pairs

Corr

Correlation; measures the correlation

etween pairs of pixels

Var

Variance; measures how spread out

the distribution of gray-levels is

Idm

Inverse Difference Moment;

measures the homogeneity

Savg Sum Average

Svar Sum Variance

Sent Sum Entropy

Ent

Entropy. Measures the randomness of

a gray-level distribution

Dvar Difference variance

Dent Difference entropy

Imcorr1 Information Measure of Corr. 1

Imcorr2 Information Measure of Corr. 2

Maxcorr Max Corr. Coefficient.

Diss Dissimilarity

Inertia Inertia

Shade Cluster Shade

Prom Cluster prominence

DSM

Digital Surface Model

(NASA SRTM 30m)

Height model, HM

DSM-DTM (Global Multi-resolution

Terrain Elevation 2010)

The feature selection phase is fundamental to

reducing the computational time of the classification

without losing accuracy (Belcore et al., 2020).

In mid-2020, the function SmileRandomForest

was introduced in the GEE coding platform. Unlike

its predecessor RandomForest function, it allows the

computation of the layer importance, which is based

on the GINI impurity system. Random Forest

algorithm creates multiple decision trees (i.e. forest)

using bootstrapped data samples and selects the final

output based on the majority vote of the individual

trees (Breiman, 2001). The Gini impurity criterion

measures the purity or randomness of a set of items

(Breiman, 2001). It determines the quality of a split

between classes in a decision tree. The goal is to split

the data in a way that results in the purest possible

subset of the classes or values. The Gini impurity of

a split is calculated as the sum of the probability of

each class or value being incorrectly classified, given

a random observation from the set. A split with a

lower Gini impurity is considered a better split, as it

results in a more pure subset of the classes or values.

The feature that results in the lowest Gini impurity is

selected as the splitting feature, and the process is

repeated until a stopping criterion is reached. The

GINI is calculated for each variable of the classifier.

The variables with high GINI gain (so they have less

impurity) are more important.

The features with less than 50 GINI gain were

removed from the input dataset as resulting of an

iterative comparison of 5 scenarios (Table 5). The

parameters considered for the best scenario

evaluation are the out-of-bag error (oob) and the

overall accuracy (OA). The threshold value was

selected according to the maximum accuracy

achievable. Finally, the features were scaled

according to their minimum-maximum values.

2.5 Classification and Multi-Temporal

Strategies Comparison

The training dataset comprises 2500 points, 300

points for each class except for the urban areas class,

which is constituted of 100 points. The choice of

unbalancing the training is due to the low percentage

of urban areas cover. Since the urban areas class

covers the smallest portion of the study area and the

Random Forest (RF) classifier tends to promote the

more represented classes in training, few samples of

the Urban areas were used to train the classifier. The

validation dataset is composed of 1800 points, 200 for

each class. The training and validation dataset were

manually created using a 2017 map as a ground

reference.

GISTAM 2023 - 9th International Conference on Geographical Information Systems Theory, Applications and Management

Two different multi-temporal approaches were

compared: aggregated multi-temporal and stacked

multi-temporal methods. In the aggregated multi-

temporal, each image was separately classified using

the machine learning algorithm Random Forest (RF)

with 100 rifle decision trees per class and 2 as the

minimum size for terminal nodes.

The same training dataset was used for each

image. The results are 16 classifications that were

aggregated according to the modal value. Only the

more accurate classifications (equal and greater than

the OA modal value) were used in the final

aggregation.

Differently, the stack multi-temporal approach

consisted of one classification over a dataset

composed of all the features from different epochs

ensembled. In this case, the images were stacked

together and classified with RF algorithm (200 rifle

decision trees per class and 4 as the minimum size for

terminal nodes). Due to GEE's limited available

memory, only 2018 and 2019 data were considered.

Figure 4 shows the classification workflow.

Figure 4: The workflow of the classification. Red arrows

indicate the DOS processing, while the blue ones indicate

the processing without DOS correction.

2.6 Accuracy Assessment

The classifications' accuracy was asses based on the

error matrix-derived measures: the overall accuracy,

the producer's accuracy, the user's accuracy, and the

F1 score (Congalton, 1991).

2.7 Comparison to Existing LC

Classifications

Today there is no official product of LC and use that

can be considered a shared and trusted reference.

Despite the large availability of satellite source data,

no high-resolution harmonised LC product exists. In

2017 ESA created a land cover classification map of

Africa at 20m resolution using 180000 Copernicus

Sentinel-2 images captured between December 2015

and December 2016 (ESA, 2016). The map is still a

prototype, and only eleven classes are described (i.e.

Trees, Shrubs, Grasslands, Croplands, Aquatic

Vegetation, Sparse Vegetation, Bare Areas, Built-Up

Snow, and Open Waters). The lack of thematic detail

is compensated by the 20m spatial resolution, which

makes it unique in the LC data of Niger. Although it

is still a partially validated prototype, it was used as

reference for the validation of the LC map generated

in this work with the stacked approach. The classes of

the two LC systems are hardly harmonised, thus, the

translation required the creation of a target shared

classification for LC Africa ESA and LC of Sirba

area, as Table 4 shows.

Table 4: Conversion classes between ESA Africa LC and

here developed LC map.

Common classes ESA Africa LC Present LC

1- Vegetation

1- Trees, 2 - Shrubs,

6 - Sparse vegetation

3 - Forest and

bushes, 8- Plateaux

vegetation

2- Grassland 3 - Grasslands

10 – non-irrigated

agricultural lands

and pastures

3- Cropland 4 - Cropland

9 – Irrigated

agricultural lands

4 - Bare areas 7- Bare areas

5 – Red bare soils,

6 - Sandy soils,

2 - Plateaux

5 - Built-up 8- Built-up 4 – Urban areas

6 - Waters

5 Aquatic vegetation,

10 - Open waters

1 – Water bodies

3 RESULTS

3.1 Feature Extraction and Selection

Figure 5 shows the results of the GINI importance

analysis. The bands per image (initially 44) were

Sentinel 2 High-Resolution Land Cover Mapping in Sub-Saharan Africa with Google Earth Engine

reduced to 16 according to the maximum achievable

accuracy (Table 5) computed by considering five

scenarios with reduced input features. Scenario 3

revealed better OA (0.85) and a little out-of-bag error

(0.070).

Table 5: Tests run over five scenarios that differ in the

number of input features in the classification selected

according to their importance value (see Figure 5).

GINI threshold oob OA

Scenario 1 none 0.081 0.846

Scenario 2 >40 0.069 0.846

Scenario 3 >50 0.070 0.854

Scenario 4 >60 0.084 0.849

Scenario 5 >80 0.081 0.842

3.2 Classification and Multi-Temporal

Strategies Comparison

The aggregated multi-temporal classification was

performed separately in 16 images with 16 features

for each. Table 6 provides the OA values calculated

for each classification. Classifications with OA less

than 0.94 were not used for the modal aggregation.

The stacked multi-temporal classification was

realised using as the input dataset the features from

different epochs together.

The period considered was 2018 and 2019. The

input features of the stack multi-temporal

classification were 76.

Figure 5: GINI importance (y) of the extracted features (x).

The GINI importance was computed for the 76

bands to optimise the process further. Four scenarios

for the slimming out were considered (Table 7). Still,

the best results in terms of OA are provided by

scenario number 1, which does not remove any

feature from the classification.

Table 6: OA achieved on each classification. The underlined

classifications were excluded from the aggregation.

Classification no. OA

1 0.97

2 0.95

3 0.96

4 0.85

5 0.94

6 0.83

7 0.85

8 0.92

9 0.95

10 0.95

11 0.92

12 0.95

13 0.95

14 0.94

15 0.95

16 0.96

Table 7: Tests run over 5 scenarios that differ in the number

of input features in the classification selected according to

their importance value. The parameters considered for the

best scenario evaluation are the out-of-bag error (oob) and

the overall accuracy (OA).

GINI threshold oob OA

Scenario 1 none 0.019 0.960

Scenario 2 >9 0.020 0.958

Scenario 3 >10 0.020 0.955

Scenario 4 >20 0.023 0.951

The atmospheric correction's influence on the

classifications was checked by comparing the

accuracy of the same classification model applied to

corrected and non-corrected input datasets. The DOS

has little impact on the aggregated multi-temporal

classification's goodness: it shifts the OA from 0.971

(non-corrected dataset) to 0.975 (corrected dataset).

Similarly, the DOS showed little influence on the

stacked method too. Indeed it shifts the OA from

0.955 (non-corrected) to 0.960 (corrected), Table 8.

Table 8: DOS influences over the classifications.

No DOS DOS

OA of Aggregated multi-temporal 0.971 0.975

OA of Stack multi-temporal 0.955 0.960

3.3 Accuracy Assessment

The error matrices give the performance of the

classification. Both multi-temporal approaches

resulted in high accuracy values. For what concerns

the aggregated approach, Table 9 shows that the

User's and Producer's accuracies are always above

GISTAM 2023 - 9th International Conference on Geographical Information Systems Theory, Applications and Management

0.95. The Plateaux class is less accurate, although its

F1 score reaches 0.95. The model correctly identifies

sandy bare soils and irrigated land classes.

Table 9: Error matrix of the aggregated multi-temporal

classification.

Water

Plateaux

Forest bushes

Urban areas

Red bare soils

Sandy bare soil

Vegetation

lateaux

)

Irrigated

ricultural lands

Non-irrigated

lands and

astures

Water 200 0 0 0 0 0 0 0 0

Plateaux 0 189 0 0 0 0 11 0 0

Forest bushes 0 0 200 0 0 0 0 0 0

Urban areas 0 0 7 193 0 0 0 0 0

Red bare soils 0 8 0 0 188 0 0 0 4

Sandy bare soil 0 0 0 0 1 197 0 0 2

Vegetation

(plateaux)

0 0 0 0 0 0

200 0 0

Irrigated

agricultural

lands

0 0 0 0 0 0 0

200 0

Non-irrigated

lands and

astures

0 0 6 5 1 0 0 0

188

TOT 200 197 213 198 190 197 211 200 194

PA 1.00 0.96 0.94 0.98 0.99 1.00 0.94 1.00 0.97

UA 1.00 0.94 1.00 0.96 0.94 0.98 1.00 1.00 0.94

F1 1.00 0.95 0.97 0.97 0.96 0.99 0.97 1.00 0.95

OA= 0.98

Table 10: Error matrix of the stacked multi-temporal

classification.

Water

Plateaux

Forest bushes

Urban areas

Red bare soils

Sandy bare soil

Vegetation

lateaux

)

Irrigated

ricultural

Non-irrigated

lands and

Water 183 0 0 0 0 0 0 0 0

Plateaux 0 180 0 0 0 0 10 0 0

Forest bushes 1 0 191 0 0 0 1 3 0

Urban areas 0 0 6 161 0 0 0 0 0

Red bare soils 0 9 0 2 174 0 0 0 8

Sandy bare soil 4 0 0 0 0 178 0 0 1

Vegetation

lateaux

)

0 1 0 0 0 0

195 0 0

Irrigated

agricultural

lands

3 0 6 0 0 0 0

171 0

Non-irrigated

lands and

astures

0 0 8 4 0 0 0 0

177

TOT 191 190 211 167 174 178 206 174 186

PA 0.96 0.95 0.91 0.96 1.00 1.00 0.95 0.98 0.95

UA 1.00 0.95 0.97 0.96 0.90 0.97 0.99 0.95 0.94

F1 0.98 0.95 0.94 0.96 0.95 0.99 0.97 0.97 0.94

OA= 0.96

Regarding the stacked classification, the accuracy

values are slightly lower than the ones on the

aggregate multi-temporal classification. Table 10

shows that the plateaux class reaches 0.947 of the F1-

score, which is the less accurate class along the non-

irrigated lands and pastures. The overall accuracy is

0.96, with only 0.05 points of difference from the

aggregated methods.

Although the high accuracy value, some salt-and-

pepper effect is present all over the scene; thus, some

post-processing operations were carried out for

aesthetic reasons. Specifically, erosion (size 4) and

dilation (size 3) were realised in class Urban areas

(Figure 6).

Figure 6: Example of the aggregated multi-temporal

classification (left) and the stacked multi-temporal

classification (right).

3.4 Comparison to Existing LC

Classifications

The pixel by pixel comparison reveals an overall

accuracy of 0.203 and the pixel-based confusion

matrix is described in Table 11. From a visual

Table 11: Error matrix of the ESA LC Africa (reference)

and Sirba Classification. To facilitate the reading, the

values are reported in square kilometres.

Vegetation

Grassland

Cropland

Bare areas

Built-up

Waters

Total

Vegetation 1068 1497 101 1568 56 97 4387

Grassland 2451 8908 105 34992 1103 787 48348

Cropland 7819 29140 1503 10711 1792 722 51686

Bare areas 36 1407 2 11037 19 113 12614

Built-up 58 4 66 59 584 12 783

Waters 10 0 101 4 0 1109 1224

Total 11442 40956 1879 58369 3554 2841

0.203

PA 0.093 0.218 0.800 0.189 0.164 0.390

UA 0.244 0.184 0.029 0.875 0.745 0.906

F1 0.135 0.200 0.056 0.311 0.269 0.546

Sentinel 2 High-Resolution Land Cover Mapping in Sub-Saharan Africa with Google Earth Engine

interpretation of the results, the water class

(specifically the Sirba river) is better identified by the

aggregated LC than the ESA LC (Figure 7).

Figure 7: Sirba River area classified according Sirba LC

(top) and ESA LC (bottom).

4 DISCUSSION

Both classification methods show very positive

results. The little influence of DOS on the results

might be caused by the short span and the very similar

meteorological condition of the analysed dataset.

Also, the RF, which is slightly sensitive to non-

normalised datasets, might contribute to such results.

Although DOS has little influence on the results, it

was maintained in the classification workflow

because of its lightweight processing time. More

complex atmospheric correction models can require

more computational power and processing time.

Thus, further and more detailed analysis needs to be

realised in this direction.

The GINI importance analysis allows the

lightening of the classification process and improves

the classification's performance. Although this was

not true for the stacked classification, reducing the

dataset reduces the OA. It is worth underlining that

the importance analysis was applied twice in this

case. The DOS correction has little influence on the

final accuracy results for both multi-temporal

approaches. This is an unexpected result since most

relevant literature underlines the importance of

atmospheric correction in multi-temporal approaches,

especially in stacked ones.

Little distance also emerges from the comparison

of the two multi-temporal approaches. The

aggregated multi-temporal classification overcomes

the stacked one for only 0.015 of OA (regardless of

the atmospheric correction). The F1 score of some

classes of the aggregated multi-temporal approach is

1 (water and irrigated lands). In the stacked multi-

temporal approaches, the F1 score shows some

differences: irrigated land class is not one of the most

accurate classes, but bare soil is. This method seems

to misclassify the irrigated class, which is often

confused with forest or water. Despite some little

differences, in this case, the two approaches are

perfectly exchangeable for this specific application.

Again, there is a little difference in terms of time to

apply one or the other.

Nevertheless, there is a strong possibility of

running out of memory in additional features or a vast

area. Indeed data from 2017 were taken out. In this

specific application, the aggregated multi-temporal

classification method was applied because of the

slightly higher accuracy and the less scarcity of salt-

and-pepper effect all over the scene.

The pixel-by-pixel comparison between the LC

ESA map and the aggregated LC reveals low overall

accuracy (0.203) despite the two classifications

having similar spatial resolutions (10m and 20m).

The primary issue regards the confusion between

Grassland, Cropland and Bare areas. Most pixels

classified as Grassland in Sirba LC are considered

Bare areas in the ESA Africa LC (Table 11).

Similarly, most of the croplands of the Sirba LC are

classified as Grassland in the ESA LC. Indeed the F1

score of Cropland is only 0.056. Such results are

ascribable to the nature of the definitions of pastures,

grasslands and bare soils. In fact, pastures are

considered Agricultural land in ESA LC and

Grassland in Sirba LC. This is clearly detectable from

the visual comparison between the classifications

(Figure 7). Sirba River and most seasonal ponds and

lakes are detected in Sirba LC and not in ESA LC

because of the dataset of the classifications. Sirba LC

is a rainy season LC (only summer months in 2017-

2019), while ESA LC is based on one-year

observations. This also influences the vegetation

class, which is captured at its maximum during the

rainy season. A good overlap is present between the

other classes. It is worth underling that the ESA

Africa LC is a prototype, and it was validated using

Crowdsourcing only for Kenya, Gabon, Ivory Coast

and South Africa. The analysis and the methodology

applied for the classification demonstrate that the

textural information facilitates segmentation and

classification.

GISTAM 2023 - 9th International Conference on Geographical Information Systems Theory, Applications and Management

Similarly, the aggregated multi-temporal approach

proposed to reduce the variability of the images led to

high-accuracy classification. Selecting a limited

period for the satellite classification allowed the

maximisation of the seasonal characterisation. It

increased the separability of some hard-to-map

classes (e.g. Nigerienne urban areas from bare soils

and pastures and water).

5 CONCLUSIONS

Regardless of the application of atmospheric

correction, the classification provides a suitable LC

map for flood planning. It follows that, with some

specific actions, it is possible to overcome the main

mapping difficulties and obtain LC maps with high

thematic detail in sub-Saharan areas. The model

proposed in this paper can be applied to classify other

sub-Saharan river areas semi-automatically since it is

developed in GEE.

REFERENCES

Anadia 2.0. climateservices.it CNR-IBE. URL https://clima

teservices.it/progetto/anadia/ (accessed 2.8.23).

Belcore, E., Piras, M., Pezzoli, A., 2022. Land Cover

Classification from Very High-Resolution UAS Data

for Flood Risk Mapping. Sensors 22, 5622. https://

doi.org/10.3390/s22155622

Belcore, E., Piras, M., Wozniak, E., 2020. Specific alpine

environment land cover classification methodology:

Google Earth Engine processing for Sentinel-2 data, in:

Volume XLIII-B3-2020. Copernicus Publications, pp.

663–670. https://doi.org/10.5194/isprs-archives-XLIII-

B3-2020-663-2020

Bigi, V., Pezzoli, A., Rosso, M., 2018. Past and Future

Precipitation Trend Analysis for the City of Niamey

(Niger): An Overview. Climate 6, 73. https://doi.org/

10.3390/cli6030073

Breiman, L., 2001. Random Forests. Machine Learning 45,

5–32. https://doi.org/10.1023/A:1010933404324

Chavez, P.S., 1988. An improved dark-object subtraction

technique for atmospheric scattering correction of

multispectral data. Remote Sensing of Environment 24,

459–479.https://doi.org/10.1016/0034-4257(88)90019-3

Congalton, R.G., 1991. A review of assessing the accuracy

of classifications of remotely sensed data. Remote

Sensing of Environment 37, 35–46. https://doi.org/

10.1016/0034-4257(91)90048-B

Conners, R. W., Trivedi, M.M., Harlow, C.A., 1984.

Segmentation of a high-resolution urban scene using

texture operators. Computer Vision, Graphics, and

Image Processing 25, 273–310. https://doi.org/

10.1016/0734-189X(84)90197-X

Dorren, L.K.A., Maier, B., Seijmonsbergen, A.C., 2003.

Improved Landsat-based forest mapping in steep

mountainous terrain using object-based classification.

Forest Ecology and Management 183, 31–46.

https://doi.org/10.1016/S0378-1127(03)00113-0

Drzewiecki, W., Wawrzaszek, A., Aleksandrowicz, S.,

Krupiński, M., Bernat, K., 2013. Comparison of

selected textural features as global content-based

descriptors of VHR satellite image, in: 2013 IEEE

International Geoscience and Remote Sensing

Symposium - IGARSS. Presented at the 2013 IEEE

International Geoscience and Remote Sensing

Symposium - IGARSS, pp. 4364–4366. https://doi.org/

10.1109/IGARSS.2013.6723801

ESA, 2016. ESA CCI LAND COVER – S2 prototype Land

Cover 20m map of Africa 2016 [WWW Document].

URL http://2016africalandcover20m.esrin.esa.int/

(accessed 4.12.21).

GEE, n.d. Google Earth Engine [WWW Document]. URL

https://earthengine.google.com/platform/ (accessed

3.9.20).

Haralick, R.M., Shanmugam, K., Dinstein, I., 1973.

Textural Features for Image Classification. IEEE

Transactions on Systems, Man, and Cybernetics SMC-

3, 610–621. https://doi.org/10.1109/TSMC.1973.

4309314

Huete, A. R., 1988. A soil-adjusted vegetation index

(SAVI). Remote Sensing of Environment 25, 295–309.

https://doi.org/10.1016/0034-4257(88)90106-X

Kukawska, E., Lewiński, S., Krupiński, M., Malinowski,

R., Nowakowski, A., Rybicki, M., Kotarba, A., 2017.

Multitemporal Sentinel-2 data - remarks and

observations, in: 2017 9th International Workshop on

the Analysis of Multitemporal Remote Sensing Images

(MultiTemp). Presented at the 2017 9th International

Workshop on the Analysis of Multitemporal Remote

Sensing Images (MultiTemp), pp. 1–4. https://doi.org/

10.1109/Multi-Temp.2017.8035212

Lantzanakis, G., Mitraka, Z., Chrysoulakis, N., 2017.

Comparison of Physically and Image Based

Atmospheric Correction Methods for Sentinel-2

Satellite Imagery, in: Karacostas, T., Bais, A., Nastos,

P.T. (Eds.), Perspectives on Atmospheric Sciences,

Springer Atmospheric Sciences. Springer International

Publishing, pp. 255–261.

Li, Q., Qiu, C., Ma, L., Schmitt, M., Zhu, X.X., 2020.

Mapping the Land Cover of Africa at 10 m Resolution

from Multi-Source Remote Sensing Data with Google

Earth Engine. Remote Sensing 12, 602. https://doi.

org/10.3390/rs12040602

Massazza, G., Tamagnone, P., Pezzoli, A., Housseini, M.,

Belcore, E., Tiepolo, M., Rosso, M., 2018. Amélio-

rations sur le système d’observation du bassin de la

Rivière Sirba pour la gestion des risques naturels, in:

Colloque International AMMA-CATCH. AMMA

CATCH.

Massazza, G., Tamagnone, P., Wilcox, C., Belcore, E.,

Pezzoli, A., Vischel, T., Panthou, G., Housseini

Ibrahim, M., Tiepolo, M., Tarchiani, V., Rosso, M.,

2019. Flood Hazard Scenarios of the Sirba River

Sentinel 2 High-Resolution Land Cover Mapping in Sub-Saharan Africa with Google Earth Engine

(Niger): Evaluation of the Hazard Thresholds and

Flooding Areas. Water 11, 1018. https://doi.org/10.

3390/w11051018

Oguntunde, P.G., Lischeid, G., Abiodun, B.J., 2018.

Impacts of climate variability and change on drought

characteristics in the Niger River Basin, West Africa.

Stoch Environ Res Risk Assess 32, 1017–1034.

https://doi.org/10.1007/s00477-017-1484-y

Poortinga, A., Tenneson, K., Shapiro, A., Nquyen, Q., San

Aung, K., Chishtie, F., Saah, D., 2019. Mapping

Plantations in Myanmar by Fusing Landsat-8, Sentinel-

2 and Sentinel-1 Data along with Systematic Error

Quantification. Remote Sensing 11, 831. https://doi

.org/10.3390/rs11070831

Shepherd, J. D., Dymond, J.R., 2010. Correcting satellite

imagery for the variance of reflectance and illumination

with topography. International Journal of Remote

Sensing. https://doi.org/10.1080/01431160210154029

Tamagnone, P., Massazza, G., Pezzoli, A., Rosso, M.,

2019. Hydrology of the Sirba River: Updating and

Analysis of Discharge Time Series. Water 11, 156.

https://doi.org/10.3390/w11010156

Teodoro, A.C., Duarte, L., 2022. Chapter 10 - The role of

satellite remote sensing in natural disaster management,

in: Denizli, A., Alencar, M.S., Nguyen, T.A., Motaung,

D.E. (Eds.), Nanotechnology-Based Smart Remote

Sensing Networks for Disaster Prevention, Micro and

Nano Technologies. Elsevier, pp. 189–216.

https://doi.org/10.1016/B978-0-323-91166-5.00015-X

Tiepolo, M., Bacci, M., Braccio, S., 2018. Multihazard Risk

Assessment for Planning with Climate in the Dosso

Region, Niger. Climate 6, 67. https://doi.org/10.3

390/cli6030067

GISTAM 2023 - 9th International Conference on Geographical Information Systems Theory, Applications and Management