论文信息

Nyatabori Kwame / Zakaria Ngereja

Journal of the Geographical Association of Tanzania

论文解读

AI 辅助整理
方法
多时相Landsat影像与生物物理、社会经济驱动因子,构建CA-Markov和R
发现
RF-CA精度77.84%、Kappa 0.67,优于CA-Markov,更好刻画城郊空间复杂性。
机会点
可将RF-CA混合模型迁移至耕地扩张或农业用地变化预测,弥补数据稀缺地区研究空白。

论文摘要

快速增长的非洲城市的城市扩张需要准确的预测工具来指导可持续城市规划。本研究比较了两种空间建模方法——元胞自动机-马尔可夫链(Cellular Automata-Markov Chain,CA-Markov)和随机森林-元胞自动机(Random Forest-Cellular Automata,RF-CA)——在预测坦桑尼亚达累斯萨拉姆城市蔓延方面的表现。研究利用多时相Landsat影像(1995年、2004年、2014年和2024年)以及关键生物物理和社会经济驱动因素,对历史动态进行建模并预测2024年的城市范围。随机森林-元胞自动机(RF-CA)模型将机器学习导出的转移概率与基于邻域的CA框架相结合,而CA-Markov模型则依赖于根据先前时间步校准的历史转移概率。结果表明,RF-CA取得了更高的预测性能(总体精度:77.84%;Kappa系数:0.67;品质因数(Figure of Merit,FoM):56.83%),优于CA-Markov(总体精度:69.38%;Kappa系数:0.542;品质因数(FoM):51.33%),并且更准确地捕捉了城郊地区的空间复杂性。人口密度、海拔和气候是最具影响力的预测因子。这些发现凸显了混合机器学习模型在快速转型城市城市增长预测中的价值,并展示了其相对于传统基于马尔可夫模型的优势。此外,本研究通过提供RF-CA在数据稀缺的非洲环境中有效性的经验证据,推进了地理信息系统与遥感(GIS and RS)领域的发展。

全文

Journal of the Geographical Association of Tanzania

Journal of the Geographical Association of Tanzania, Vol. 46, No. 2: 1–27
© Department of Geography, UDSM, 2026 ▪ https://doi.org/10.56279/jgat.v46i2.389
A Comparative Evaluation of the CA-Markov and Random
Forest-Cellular Automata Models for Predicting Urban
Sprawl in Dar es Salaam, Tanzania

Nyatabori Antonietta Kwame* & Zakaria Ngereja§

Abstract
Urban expansion is occurring rapidly in African cities, necessitating accurate
predictive tools to guide sustainable urban planning. This paper draws data
from a study that compared the performance of two spatial modelling
approaches—Cellular Automata -Markov Chain (CA -Markov) and Random
Forest-Cellular Automata (RF -CA)—in forecasting urban sprawl in Dar es
Salaam, Tanzania. Multitemporal Landsat imagery (1995, 2004, 2014, and
2024) and key biophysical and socioeconomic drivers were used to model the
past dynamics and predict the 2024 urban extent. The Random Forest-Cellular
Automata (RF -CA) model integrates machine -learning-derived transition
probabilities with a neighbourhood-based CA framework. In contrast, the CA-
Markov model relies on historical transition probabilities calibrated from
previous time steps. The re sults showed that RF -CA achieved a higher
predictive performance (accuracy: 77.84%; Kappa: 0.67; Figure of Merit (FoM):
56.83%) than CA -Markov (accuracy: 69.38%; Kappa: 0.542; Figure of Merit
(FoM): 51.33%), and more accurately captured the spatial complexity in peri-
urban areas. These results imply that RF -CA is a more reliable model for
predicting urban growth, especially in complex, rapidly changing peri -urban
areas. Population density, elevation, and clim ate were found as the most
influential factors in this study. These results highlight the value of hybrid
machine-learning models for urban growth prediction in rapidly transforming
cities and demonstrate their advantages over traditional Markov -based
models. Furthermore, this paper generates empirical evidence on the
applicability of RF -CA for urban growth modelling in data -scarce A frican
contexts using geospatial and remote sensing methods.
Keywords: Urban sprawl, CA-Markov, Random Forest - Cellular
Automata, GIS and Remote Sensing, Sub-Saharan Africa

  1. Introduction
    Urban sprawl has accelerated globally, driven by population growth, economic
    development, and increasing land demands. Between 1990 and 2014, global
    urban sprawl increased by approximately 95% ; equivalent to 4% per year
    (Behnisch, Krüger & Jaeger, 2022) . This rapid expansion has led to the
    conversion of agricultural land, proliferation of informal settlements, and
    heightened pressure on infrastructure, particularly in developing countries

*Department of Geospatial Sciences and Technology, Ardhi University, Dar es Salaam, Tanzania :
nyataborikwame@gmail.com (Corresponding author)
§ Department of Geospatial Sciences and Technology, Ardhi University, Dar es Salaam, Tanzania

Nyatabori Antonietta Kwame & Zakaria Ngereja
2 JGAT Volume 46, Number 2, 2026
where planning systems often lag behind urbanisation (Seto, Güneralp &
Hutyra, 2012) . Therefore, accurately predicting future urban expansion
patterns is essential for sustainable urban planning and management (Mnyali
& Materu, 2021; Wang, Shi & Zhou, 2020).

Dar es Salaam, Tanzania’s largest city and primary economic hub, exemplifies
these challenges as one of Africa’s fastest-growing cities. Its population surged
from 4.3 million in 2012 to 5.3 million in 2022 (URT, 2022), thereby fuelling
complex spatial dynamics characterised by informal settlement patterns and
coastal vulnerability (Batty, Xie & Sun, 1999; Augustijn -Beckers, Flacke &
Retsios, 2011).

Traditional land-use prediction methods struggle to capture heterogeneous and
rapidly evolving land -use patterns (Jayabaskaran & Das, 2023) . However,
advances in geospatial analysis and machine learning have enhanced urban
growth modelling. Cellular Automata (CA) integrated with a Markov Chain –
known as CA-Markov is a widely used approach that combines historical
transition probabilities with neighbourhood rules to simulate land-use changes
(Sang et al., 2011; Hamad, Balzter & Kolo, 2018; Simon, Lyimo & Yamungu,
2023; Atef, Ahmed & Abdel-Maguid, 2024). However, the CA -Markov model
assumes stationary historical transitions, thereby limiting its ability to
represent dynamic nonlinear interactions in rapidly urbanising environments
(Nugroho & Al-Sanjary, 2018).

To overcome these limitations, hybrid models incorporating machine
learning—such as Cellular Automata (CA) and Artificial Neural Networks
(ANNs) (Gharaibeh et al., 2020) and random forest-cellular automata (RF-CA)
(Kamusoko & Gamba, 2015) —have emerged. The Random Forest —an
ensemble method (Breiman, 2001)—excels at handling high-dimensional data,
feature selection, and nonlinear relationships. When coupled with CA, RF-CA
generates spatially explicit predictions by deriving transition probabilities
from machine learning while retaining neighbourhood effects (Gharaibeh et al.,
2020; Khawaldah, Farhan and Alzboun, 2020; Asif et al., 2023). Kamusoko and
Gamba (2015) demonstrated the superiority of RF-CA over the Support Vector
Machine Cellular Automata (SVM -CA) and the Logistic Regression Cellular
Automata (LR-CA) in capturing heterogeneous urban patterns.

Nonetheless, there is limited empirical evidence comparing the predictive
performance of RF-CA with the CA-Markov approach in rapidly urbanising cities
of Sub-Saharan Africa despite these advances (Laros et al., 2014; Mberu, Béguy
and Ezeh, 2017; Tumwesigye et al., 2023). Most RF-CA evaluations have focused
on Asia and Europe, leaving a gap in understanding its performance in data-scarce
and informally driven African contexts, such as Dar es Salaam.

CA-Markov and Random Forest-Cellular Automata Models for Predicting Urban Sprawl

3 JGAT Volume 46, Number 2, 2026
Addressing this gap requires not only applying RF -CA in an African context,
but also benchmarking it against the traditional CA-Markov approach—which
is still widely used by planners in the region—to demonstrate whether, and by
how much , the added complexity of the machine -learning hybrid model
improves on established practice. Therefore, the research gap addressed in this
study is the limited comparative evaluation of RF -CA and CA-Markov within
the same urban context and evaluation framework.

The specific objectives were to: (i) identify the key drivers of urban expansion;
(ii) analyse spatial and temporal growth dynamics over the past three decades;
and (iii) evaluate the predictive performance of both models using multiple
accuracy metrics. By applying RF -CA to Dar es Salaam —one of East Africa ’s
fastest-growing cities, and (to our knowledge) an understudied context for this
hybrid model—this study contributes to the limited evidence base on machine-
learning integrated urban growth modelling in Sub-Saharan Africa.

  1. Theoretical Framework
    Urban sprawl does not occur as a single uniform event; it is a cumulative
    outcome of many local decisions that accumulate into identifiable area patterns.
    It is primarily driven by population growth, land demand, accessibility,
    environmental suitability, and existing built forms (Azabdaftari & Sunar, 2024;
    Fitawok & Minale, 2024; Li et al., 2025). This study draws on a complexity theory
    that frames cities as self -organising systems whose large-scale form emerges
    from simple and local rules of interaction (Spencer, 2009), which is in harmony
    with Tobler’s (1970) principle that nearby locations influence one another more
    strongly than distant ones. Cellular automata (CA) operationalise this idea by
    modelling land as a grid of cells whose states evolve through neighbourhood -
    based transition rules (White & Engelen, 1993) . Within a CA -Markov
    framework, CA supplies this spatial logic, whereas a Markov chain supplies the
    temporal logic using historically observed transition rates to project the extent
    of change likely to occur (Azabdaftari & Sunar, 2024).

However, traditional CA -Markov models are based on relatively simplistic
rules that cannot effectively characterise the complex and context -dependent
relationship between expansion and its drivers. By incorporating an ensemble
of decision trees into the ce llular automata framework, the random forest
cellular automata (RF -CA) model solves this problem by estimating change
probabilities using multiple interactive nonlinear predictors (Kamusoko &
Gamba, 2015) , given the potential of the Random Forest model to avoid
overfitting and average predictions among resampled trees (Breiman, 2001), as
well as the trend toward using machine -learning techniques in geospatial
analysis (Shaamala et al ., 2025) . In both cases, this is possible because
geospatial technologies can combine space and time into forecasts useful for

Nyatabori Antonietta Kwame & Zakaria Ngereja
4 JGAT Volume 46, Number 2, 2026
planning purposes (Musa, Hashim & Reba, 2017). Thus, comparing CA-Markov
and RF-CA is relevant to understand which of the two approaches—rule-based
or data-driven—provides the most accurate representation of urban dynamics
in rapidly growing cities, such as Dar es Salaam.

  1. Materials and Methods
    3.1 Study Area
    Dar es Salaam is the largest city and major port of Tanzania that borders the
    Indian Ocean, with an area of 1,590km2. It lies between 6°36′S and 7°0′S latitude,
    and 39°0′E and 39°33′E longitude. The city is the country’s main economic hub,
    and consists of five administrative districts: Ilala, Kinondoni, Temeke, Ubungo,
    and Kigamboni. According to the 2022 National Population and Housing Census,
    the population of Dar es Salaam increased from 4.3 million residents in 2012 to
    5.3 million residents in 2022, making it one of the fastest-growing cities in East
    Africa. The city experiences a tropical coastal climate characterised by high
    humidity and warm temperatures throughout the year, with annual rainfall
    between 1,000 and 1 ,150mm. The temperature ranges from 25°C during the
    cooler months, to 38°C during the hottest months of the year. These demographic
    and climatic conditions, combined with ongoing urbanisation and informal
    settlement expansion, make Dar es Salaam a critical case study for examining
    urban sprawl, infrastructure challenges, and environmental management.

Figure 1: Location Map of Dar es Salaam
Source: Authors

CA-Markov and Random Forest-Cellular Automata Models for Predicting Urban Sprawl

5 JGAT Volume 46, Number 2, 2026
3.2 Data Sources and Preprocessing
This study used multiple geospatial datasets ( Table 1) covering the Dar es
Salaam metropolitan area in 1995, 2004, 2014, and 2024 for the analysis.
Landsat 5 TM and Landsat 8 OLI images, with a 30m resolution, were obtained
from the USGS Earth Explorer platform for land cover classification. T he
topographical data, including elevation and slope, were derived from the Shuttle
Radar Topography Mission; and processed in QGIS using the OpenTopography
plugin. The accessibility and road network data were obtained from
OpenStreetMap (OSM), and were used to generate the Euclidean distance to a
road network. The population density data were retrieved from the Global
Human Settlement (GHS) dataset (Schiavina et al., 2023).

Table 1: Data Used for Urban Sprawl Analysis in Dar es Salaam (1995–2024)
Data Type Spatial
Resolution
Temporal
Coverage
Source Description
Satellite Imagery
(Landsat)
30m 1995, 2004,
2014 and
2024
Landsat 5 (TM), 8
(OLI) (USGS
Earth explorer)
Land Use/Land Cover
(LULC)
classification
Tanzania
Administrative
Boundaries
vector 2022 Africa Geoportal

Determining the
extent of the analysis
DEM (Digital
Elevation Model)
and slope
30m static SRTM (30m) Indicates terrain
steepness
Road Network Data Vector Static (2024) OpenStreetMap
(OSM) Geofabrik
(website)
Assessing
transportation
impacts on urban
expansion
Population
Density
1km
(resampled)
2004–2024 Global Human
Settlement (GHS)

Understanding
population-driven
urban expansion
Climate Data 5km
4km
(resampled)
2004–2024 CHIRPS
(Rainfall)
TERRA
CLIMATE
(Temperature)
Assessing climate-
related urban growth
constraints

The climatic variables were obtained directly from official data repositories,
with rainfall derived from the CHIRPS Version 2.0 daily precipitation data at
a spatial resolution of 0.05° (~5 km). On the other hand, the maximum and
minimum temperatures (Tmax and Tmin) data were obtained from the
TerraClimate monthly climate dataset at a spatial resolution of approximately
1/24° (~4km) for the study years 1995, 2004, 2014, and 2024. The datasets were
downloaded for the study area and processed by clipping them to the
boundaries of Dar es Salaam. The CHIRPS daily precipitation was aggregated
to annual rainfall totals (mm), while the monthly TerraClimate Tmax and

Nyatabori Antonietta Kwame & Zakaria Ngereja
6 JGAT Volume 46, Number 2, 2026
Tmin values were aggregated to annual temperature measures. The CHIRPS
data were accessed from the Climate Hazard Centre data and TerraClimate
repositories (Funk et al., 2015; Abatzoglou et al., 2018).

All datasets were projected to the coordinate reference system of the land cover
data, and aligned to a common spatial grid before model development using R
Studio (version 4.4.3). The 2004 land cover raster was used as the reference raster
to standardise the spatial extent, resolution, and pixel alignment of the predictor
layers using nearest-neighbour interpolation (NGB). Feature selection was based
on the relevance of the features to urban expansion: population density and
accessibility influence settlement distribution ; slope affects buildable land ;
distance to roads indicates connectivity; and rainfall and temperature reflect
environmental suitability for land conversion. This preprocessing workflow
ensured that the multitemporal datasets were compatible with model calibration,
simulation, and validation; following the methodological flowchart in Error!
Reference source not found..

The Landsat imagery comprised Collection 2 Level-2 Surface Reflectance scenes
(Landsat 5 TM for 1995 and 2004; Landsat 8 OLI for 2014 and 2024), obtained
from the USGS Earth Explorer platform (https://earthexplorer.usgs.gov), with
scenes selected from the dry season (June–September) of each target year to
minimise cloud cover. The Digital Elevation Model was the SRTM 1 Arc-Second
Global (SRTMGL1 v3, 30m) product, accessed through the OpenTopography API
(https://opentopography.org), and processed for slope derivation in QGIS. Road
network vectors were obtained from the OpenStreetMap Tanzania extract
distributed by Geofabrik (https://download.geofabrik.de). Population density was
obtained from the Global Human Settlement Population Grid ( https://human-
settlement.emergency.copernicus.eu/download). The rainfall data were sourced
from CHIRPS v2.0 (0.05° native resolution, resampled to 5km; Climate Hazards
Centre (https://data.chc.ucsb.edu/products/CHIRPS-2.0)), while the temperature
data were obtained from TerraClimate (4 km native resolution; University of
Idaho Climatology Lab (https://www.climatologylab. org/terraclimate.html)).

The land cover classification was performed using Google Earth Engine (GEE)
with a Random Forest (RF) classifier. Four land cover classes were defined:
water, vegetation, built -up land, and bare land. T he t raining samples were
generated through the visual interpretation of high -resolution basemaps in
Google Earth Pro (for example, Google Imagery) with a set of reference samples.
The RF classifier was trained using 500 trees, with 80% of the samples used for
training and 20% for validation, and with a random stratified split for each year.
The validation set comprised 120 reference points (distributed across water,
vegetation, built-up, and bare land); and the full per-year and per-class confusion
matrix are reported in the Appendix. These classified land cover maps served as

Figure 1: Methodological Flowchart for Urban Sprawl Prediction

Nyatabori Antonietta Kwame & Zakaria Ngereja
8 JGAT Volume 46, Number 2, 2026
3.3 Model Implementation
3.3.1 RF-CA Model
The RF-CA model architecture combines the Random Forest implementation
of scikit-learn with a custom Cellular Automata module in Python. The model
was implemented using Python 3.12 and JupyterLab (version 3.6.3 ). The
Random Forest classifier was trained using predictor variables derived from
the initial land cover state of 2004 –2014. The predictor set consisted of ten
variables: land cover, Moore neighbourhood, extended neighbourhood, slope,
distance to roads, population density, DEM, mean temperature, rainfall, and
distance to urban areas. Predictor variables were used in their native units
without normalisation or rescaling since the Random Forest partitions are
based on the relative ordering of feature values at each split rather than the
absolute distance between observations. The dataset was divided into 70%
training and 30% testing subsets using a random fixed seed of 42.

To address class imbalance, all observed urban-transition cells were retained,
whereas non-transition non-urban cells were randomly sampled at a maximum
ratio of 3:1. The resulting dataset was divided into 70% training and 30%
testing subsets using a random seed of 42. Unlike the land cover classification
Random Forest in Section 3.2, the RF -CA transition model was trained on a
balanced sample of pixels using a 70/30 training-testing split.

To minimise temporal information leakage, predictor variables were derived
from the initial state of each historical calibration interval, whereas the
subsequent land cover was used only to construct the transition target. For the
2004–2014 calibration, predictor information from 2014 was not used to train
the Random Forest, but was used only to define the binary transition target.

For a given cell 𝑖, the probability of transitioning to urban class 𝑃𝑖,𝑡 was
calculated as the mean prediction across all trees in a forest:
𝑃𝑖,𝑡 = 1𝑁𝑡𝑟𝑒𝑒𝑠 ∑ 𝐼
𝑁𝑡𝑟𝑒𝑒𝑠
𝑘=1
(ℎ𝑘(𝑋𝑖,𝑡)) (1)
Where:
𝑃𝑖,𝑡 = predicted probability of urban transition for pixel 𝑖 at time 𝑡
ℎ𝑘(𝑋𝑖,𝑡) = prediction of the k-th decision tree for the feature vector 𝑋𝑖,𝑡 of
pixel 𝑖
𝑁𝑡𝑟𝑒𝑒𝑠 = total number of trees in the forest.
𝐼 = indicator function (1 if the tree predicts “urban”, 0 otherwise)
𝑋𝑖,𝑡 = [ 𝐿𝐶𝑖.𝑡, 𝑁𝑖,𝑡(3×3), 𝑁𝑖,𝑡(5×5), 𝑆𝑖, 𝑅𝑖, 𝐷𝐸𝑀𝑖, 𝑃𝑂𝑃𝑖,𝑡, 𝑅𝑎𝑖𝑛𝑖,𝑡, 𝐷𝑢𝑖,𝑡, 𝑇𝑖,𝑡]
Where:

CA-Markov and Random Forest-Cellular Automata Models for Predicting Urban Sprawl

9 JGAT Volume 46, Number 2, 2026
𝐿𝐶𝑖.𝑡 = land cover class at pixel 𝑖 at time 𝑡
𝑁𝑖,𝑡(3×3) = number of urban cells within the 3×3 neighbourhood
surrounding pixel 𝑖 at time 𝑡 (Moore neighbourhood)
𝑁𝑖,𝑡(5×5) = number of urban cells within the 5×5 neighbourhood
surrounding pixel 𝑖 at time 𝑡 (extended neighbourhood)
𝑆𝑖 = slope at pixel 𝑖
𝑅𝑖 = distance from pixel 𝑖 to the major road
𝐷𝐸𝑀𝑖 = digital elevation at pixel 𝑖
𝑃𝑂𝑃𝑖,𝑡 = population density at pixel 𝑖 at time 𝑡
𝑅𝑎𝑖𝑛𝑖,𝑡 = Rainfall at pixel 𝑖 at time 𝑡
𝐷𝑢𝑖,𝑡 = distance from pixel 𝑖 to existing urban areas at time 𝑡
𝑇𝑖,𝑡 = mean temperature of pixel 𝑖 at time 𝑡

The resulting transition probabilities were integrated into a CA simulation
using an annual iterative allocation procedure. The CA simulation was run
iteratively at an annual time -step for ten iterations. At each iteration, 3 ×3
Moore and 5 ×5 extended windows were recalculated from the updated land
cover state to represent the local and broader spatial influences of existing and
newly urbanised cells, respectively. The formula used was:

𝑆𝑡 = 𝑇𝑜𝑝𝑁𝑡(𝑃𝑖,𝑡) (2)
Where:
𝑆𝑡 = set of pixels selected for conversion to the urban class at iteration 𝑡
𝑇𝑜𝑝𝑁𝑡 = selection of the 𝑁𝑡 cells with the highest predicted transition
probabilities
𝑃𝑖,𝑡 = predicted probability of urban transition
𝑁𝑡 = number of cells to be converted during iteration 𝑡

The RF transition probabilities 𝑃𝑖,𝑡 were passed to the CA allocation procedure,
which ranks all non -urban 𝑁𝑡 cells, and selects the cells with the highest
probabilities as per the formula:
𝑁𝑡 = [0.01𝑁𝑛𝑢,𝑡] (3)
Where:
𝑁𝑡 = number of cells to be converted during iteration 𝑡
0.01 = annual transition rate (1%)
𝑁𝑛𝑢,𝑡 = number of non-urban cells

The number of cells converted during each iteration was determined using a
fixed transition rate of 1% of the currently non -urban cells. After each

Nyatabori Antonietta Kwame & Zakaria Ngereja
10 JGAT Volume 46, Number 2, 2026
allocation step, the land cover was updated, and the 3 ×3 and 5 ×5
neighbourhood measures were recalculated annually until the specific target
year was reached.

The simulated 2024 urban extent was compared with an independently classified
2024 land cover map derived from Landsat satellite images. For the 2024–2034
simulation, the 2024 predictor layer was held constant throughout the CA
iteration, while the land cover state and neighbourhood windows were updated
annually. Subsequently, the 2034 RF -CA output represented a conditional
projection of urban expansion based on the 2024 predictor variables.

3.3.2 CA-Markov Model
For the CA-Markov implementation in TerrSet (v19.0.6), the transition potential
analysis focused on the 2004–2014 period as the calibration interval, with the CA
simulation run iteratively at an annual time step of 10 iterations per period to
capture yearly urban growth dynamics, excluding earlier transitions (1995– 2004).

According to Wang, Munkhnasan and Lee (2021) and Simon, Lyimo and
Yamungu (2023), the Markov Chain model calculates a transition probability
matrix where each element 𝑃𝑖𝑗 represents the probability that a pixel of class
𝑖 will change to class 𝑗 over a specified time interval, as follows:
𝑃𝑖𝑗 = [
𝑃11 𝑃12
𝑃21 𝑃22 ⋯ 𝑃1𝑛⋯ 𝑃2𝑛
⋮ ⋮
𝑃𝑛1 𝑃𝑛2 ⋱ ⋯⋯ 𝑃𝑛𝑚] (4)

The model was then calculated as follows:
𝐿𝑆(𝑡+1) = 𝑃𝑖𝑗 × 𝐿𝑆𝑡 (5)
Where:
𝐿𝑆(𝑡+1) represents Land Cover at time (𝑡 + 1)
𝐿𝑆𝑡 represents Land Cover at time 𝑡

The Markov chain model used the 2004 Land Cover as the starting image and
the 2014 Land Cover as the ending image. The transition probability matrix
was used alongside a set of suitability maps derived from driving factors—such
as distance to roads, slope, elevation, and population density—using the Multi-
Criteria Evaluation (MCE) module in TerrSet ; and was then used in the CA
simulation to predict the 2024 urban extent. For future projections, the
transition probabilities from 2014 to 2024 were used to project the conditional
2034 scenario, while the same MCE-derived suitability maps and weights were
held constant.

CA-Markov and Random Forest-Cellular Automata Models for Predicting Urban Sprawl

11 JGAT Volume 46, Number 2, 2026
3.4 Parameter Calibration and Selection
3.4.1 CA-Markov Model
The CA-Markov model was parameterised in TerrSet (v19.0.6) using a 10-year
calibration/prediction interval, represented by ten annual CA iterations. The
2004–2014 land-cover transition was used for model calibration and validation
of the 2024 urban extent, and the 2014–2024 transition was subsequently used
to derive the transition information for the 2034 scenario. Transition quantities
were estimated using the Markov Chain component, and spatial allocation was
controlled using suitability maps generated through Multi-Criteria Evaluation
based on distance to roads, slope, elevation, and population density. The CA
component then spatially allocates the projected transitions using the
configured neighbourhood and suitability information.

3.4.2 RF-CA Model
A 3×3 Moore neighbourhood and a 5×5 extended neighbourhood measure were
recalculated at each CA iteration and supplied as separate predictors to the
Random Forest model. The Random Forest consisted of 100 trees with a fixed
random seed of 42, and the transition dataset used a 70/30 training -testing
split. Class imbalance was addressed as follows: all non -urban cells were
sampled at a maximum ratio of 3:1. The CA component used an annual time
step and a transition rate of 1%, whereby eligible non-urban cells were ranked
according to their RF -derived transition probabilities, and the cells with the
highest probabilities were converted to the urban class.

Table 2: RF-CA Neighbourhood Window Comparison
Configuration Cells Approx.
coverage
Role in this study
3×3 8 neighbours 90m×90m Local neighbourhood predictor
5×5 24
neighbours
150m×150m Extended neighbourhood
predictor

3.5 Model Validation and Accuracy Assessment
The predictive performance of each model was evaluated by comparing the
simulated 2024 land cover maps with an independently classified reference
map for 2024. The accuracy of the model was assessed using several
complementary metrics. The overall accuracy was computed as the ratio of the
correctly classified pixels to the total number of pixels in all land cover classes.
Kappa simulation was used to measure the level of agreement between the
predicted and reference maps. To specifically evaluate the urban c lass
performance, an urban -specific F1-score was calculated to balance precision
and recall, as defined below:
𝐹1 = 2 × 𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛 × 𝑅𝑒𝑐𝑎𝑙𝑙 𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛 + 𝑅𝑒𝑐𝑎𝑙𝑙 (6)

Nyatabori Antonietta Kwame & Zakaria Ngereja
12 JGAT Volume 46, Number 2, 2026
Furthermore, the Figure of Merit (FoM) was used to assess the accuracy of the
predicted land-use change from non -urban to urban areas between 2014 and
2024. The FoM quantifies the proportion of correctly predicted changes relative
to the total observed and predicted changes, which is calculated as:
𝐹𝑜𝑀 = 𝐻𝐻 + 𝑀 + 𝐹 (7)
Where:
𝐻 represents hits (correctly predicted changes)
𝑀 denotes misses (observed changes that are not predicted)
𝐹 denotes false alarms (predicted changes that are not observed).

Only pixels that transitioned to urban areas in the reference map were included
in the FoM calculations. In addition, a 2×2 was constructed for each model, from
which the user ’s and producer ’s accuracies for the urban class were derived,
alongside quantity and allocation disagreement, following Pontius and Millones
(2011). Quantity disagreement is the absolute difference between the predicted
and observed urban areas attributable to over-/or under-estimation of the total
quantity disagreement is the area misclassified due to spatial mismatch despite
the correct overall quantity. In contrast, allocation disagreement is the area
misclassified due to spatial mismatch despite the correct overall quantity.

Spatial patterns of agreement and disagreement were further examined using
side-by-side comparisons of predicted and observed urban extents (Figures 5
and 6), while transition matrices summarised class-to-class land cover changes
between 2004 and 2014, and 2014 and 2024 (Figures 7 and 8). These analyses
provided a robust assessment of each model ’s ability to capture the spatial
complexity of urban expansion, particularly in peri-urban areas.

  1. Results
    4.1 Land Cover Classification Accuracy and Urban Expansion Trends
    Random Forest classification achieved high accuracy in distinguishing land cover
    classes across all study years, with the overall accuracies ranging from 83.33% to
    87.5%; and the kappa values ranging from 0.765 to 0.824. The full per -year
    confusion matrices, sample sizes, and user/producer accuracies are reported in the
    Appendix. This reliable classification provides a robust foundation for analysing
    urban sprawl in cities. The urban extent increased substantially from 147.40km²
    (in 1995) to 505.01 km² (in 2024), a gain of 357.62km² (Figures 3 and 4). The
    expansion occurred in three distinct phases: between 1995 and 2004 the urban
    land increased by 29.56km² (20.05%); between 2004 and 2014 the growth was
    more rapid, adding 152.20km² (86.01%); and expansion continued strongly
    between 2014 and 2024 with a 175.86km² (53.43%) expansion, marking the largest
    absolute increase in the study period. Most of this growth occurred at the expense
    of vegetation and bare land, whereas the water area remained relatively stable.

CA-Markov and Random Forest-Cellular Automata Models for Predicting Urban Sprawl

13 JGAT Volume 46, Number 2, 2026

Figure 2: Land Cover Classification (1995 to 2024) Using
the Random Forest (RF) Algorithm

Figure 3: Trends in Urban Expansion and Growth Rate
in Dar es Salaam (1995–2024)
Sources: Authors
86.01%

Nyatabori Antonietta Kwame & Zakaria Ngereja
14 JGAT Volume 46, Number 2, 2026
4.2 Model Performance Comparison
When comparing the predictive models, both CA-Markov and RF-CA captured
the general trend of outward expansion but differed in spatial detail and accuracy.
The CA-Markov model projected an urban area of 457.30km² by 2024, whereas
the RF-CA model projected 441.64km², both of which underestimate the actual
size of 505.01km². However, RF-CA produced more accurate results, correctly
predicting 343.05km² of urban area compared with 326.43km² for CA-Markov. It
also reduced errors, with lower under-prediction (161.96 vs 178.59km²) and over-
prediction (98.59 vs 130.88km²) compared with CA-Markov.

Table 3: Cross Tabulation of Correctly Predicted, Missed and Falsely
Predicted Urban Areas for CA-Markov and RF-CA Models
Model Correctly
Predicted (TP)
Missed Urban
(FN)
False
Urban (FP)
User’s/Producer’s
Accuracy
CA-Markov (km2) 326.43 178.59 130.88 71.4%/64.6%
RF-CA (km2) 343.05 161.96 98.59 77.7%/67.9%

Table 4 indicates that although RF-CA slightly underestimated the total urban
quantity by a larger margin than CA -Markov, it produced substantially less
spatial misallocation: meaning its errors were more a matter of predicting the
wrong exact location than the wrong overall amount of growth, a distinction
that the raw accuracy and Kappa alone do not reveal.

Table 4: Quantity and Allocation Disagreement

Table 5 shows that the RF-CA model consistently outperformed the CA-Markov
model across all evaluation metrics. It achieved 77.84% accuracy, Kappa 0.67, F1-
score 72.5%, and FoM 56.83%; whereas CA-Markov achieved 69.38% accuracy,
Kappa 0.542, F1-score 67.8%, and FoM 51.33%. These results demonstrate that
RF-CA not only reduces prediction errors, but also captures the spatial
complexity of urban growth more effectively than other models.

Table 5: Performance Comparison Between RF-CA and CA-Markov Model
Metric RF-CA CA-Markov
Accuracy 77.84% 69.38%
Figure of Merit (FoM) 56.83% 51.33%
Kappa Simulation 0.67 0.542
F1 Score 72.5% 67.8%
Model Quantity
Disagreement
(km2)
Allocation
Disagreement
(km2)
Total
Disagreement
(km2)
Main Errors
Characteristic
CA-Markov 47.71 261.76 309.47 High spatial misallocation
RF-CA 63.37 197.18 260.55 Higher quantity error but
lower spatial misallocation

CA-Markov and Random Forest-Cellular Automata Models for Predicting Urban Sprawl

15 JGAT Volume 46, Number 2, 2026
4.3 Spatial Patterns of Predicted Urban Growth
The RF-CA model predicted urban sprawl in Dar es Salaam for 2024 (Figure 5),
showing edge expansion and infill development mainly around existing built-up
areas such as Ubungo, Kimara, and Mbezi. Growth is influenced by population
pressure and proximity to major roads, resulting in a visually contiguous pattern
of urban growth ar ound these nodes. This description is based on a visual
inspection of the predicted map rather than quantified landscape metrics; and is,
therefore, presented as a qualitative observation only. Figure 5 shows a side-by-
side comparison of the projected (left) and actual (right) land cover in 2024.

Figure 4: Projected Land Cover versus Actual Land Cover in 2024
by Random Forest – Cellular Automata (RF-CA) model.
Source: Authors

Figure 6 shows a side-by-side comparison of projected and actual land cover.
The CA-Markov model’s 2024 prediction shows urban expansion primarily over
bare land and vegetation, driven by historical land cover transitions and
spatial factors such as slope and road distance. The expansion was
concentrated near roads and in low-slope areas, resulting in a more linear and
constrained growth pattern than the RF-CA model predictions.

Nyatabori Antonietta Kwame & Zakaria Ngereja
16 JGAT Volume 46, Number 2, 2026
Figure 5: Projected Land Cover versus Actual Land Cover
of 2024 by (CA-Markov) Model
Source: Author)

4.4 Transition Probabilities and Driving Factors
A transition probability analysis illuminated the key land conversion processes.
The 2004–2014 period, as shown in Figure 7, showed a vegetation -to-urban
transition of 3.5% , and a bare land -to-urban change of 17.4%. These rates
increased to 5.3% and 28.0%, respectively, in the 2014–2024 projection (Figure 8).

Figure 6: Land Cover Transition Matrix from 2004–2014.
Source: Authors

0.00
%
10.00
%
20.00
%
30.00
%
40.00
%
50.00
%
%
60.00
70.00
%
%
80.00
90.00
%
Water
Vegetation
Built-up
Bareland
Water
%
67.40
%
10.30
%
2.20
20.10
%
Vegetation
0.70
%
%
57.80
%
3.50
38.00
%
Built-up
0.00
%
%
0.30
%
82.90
16.70
%
Bareland
0.00
%
8.50
%
17.40
%
74.10
%

CA-Markov and Random Forest-Cellular Automata Models for Predicting Urban Sprawl

17 JGAT Volume 46, Number 2, 2026

Figure 7: Land Cover Transition Matrix from 2014–2024.
Source: Authors

In addition to its superior predictive performance, the RF -CA model provides
insights into the main drivers of sprawl through a feature importance analysis
as shown in Figure 9.

Figure 8: Feature Importance (%) Ranking for Urban Sprawl
Prediction Using the RF-CA Model
Source: Authors

0.00
10.00
%
20.00
%
30.00
%
40.00
%
%
50.00
%
60.00
70.00
%
80.00
%
Water (1)
Vegetation
(2)
Urban (3)
Bareland (4)
Water (1)
%
49.57
5.85
%
%
20.66
%
23.92
Vegetation (2)
0.97
%
%
75.14
5.34
%
%
18.55
Urban (3)
0.02
%
3.71
%
%
76.90
19.37
%
Bareland (4)
0.01
%
%
12.91
28.03
%
59.06
%

Nyatabori Antonietta Kwame & Zakaria Ngereja
18 JGAT Volume 46, Number 2, 2026
Figure 9 shows that population density (23.06%) was the dominant factor in
urban sprawl; followed by elevation (13.19%), temperature (11.75%), rainfall
(11.45%), and distance to roads (10.90%). Other contributors included slope
(10.07%), and distance to existing urban areas (8.67%). The neighbourhood
effects contributed comparatively less to the trained model, with the extended
(5×5) neighbourhood accounting for 4.95%, and the Moore (3×3) neighbourhood
accounting for 2.60%. Both neighbourhood variables were retained as separate
predictors in the RF-CA simulation (see Section 3.3).

The RF importance was assessed using impurity -based feature importance
provided by scikit-learn (feature importance), which represents the mean decrease
in impurity contributed by each predictor across the trees in the fitted forest. These
values represent model associations; and should not be interpreted as evidence of
the causal effects of population, rainfall, or any other predictor of urban growth.

4.5 Future Projections to 2034
The 2034 outputs shown in Figure 10 are conditional scenarios because neither
model incorporates future changes in population, climate, or road infrastructure.
Future projections to 2034, as shown in Figure 10, indicate continued outward
expansion under both models, but with distinct spatial patterns.

Figure 9: Land Cover 2034 showing Expansion of Urban Pattern
Using Both CA -MARKOV and RF-CA Approaches
Source: Authors

CA-Markov and Random Forest-Cellular Automata Models for Predicting Urban Sprawl

19 JGAT Volume 46, Number 2, 2026
As per Figure 10, the CA-Markov model suggested radial growth with some
leapfrog development into isolated areas. In contrast, the RF -CA model
indicated a more compact and organised expansion characterised by edge
growth and infill development along the transport corridors. Ove rall, the
results demonstrate that Dar es Salaam has undergone rapid urban expansion
over the past three decades, with machine -learning integrated approaches—
such as RF -CA—providing more accurate and spatially realistic predictions
than traditional models, such as CA-Markov.

  1. Discussion
    5.1 Model Performance and Comparative Advantages
    A comparative analysis of the CA-Markov and RF-CA models revealed significant
    differences in their predictive capabilities for forecasting urban sprawl in Dar es
    Salaam, Tanzania. The results indicate that the hybrid RF -CA model
    outperformed the traditional CA -Markov model, achieving higher accuracy
    (77.84% vs 69.38%), Figure of Merit (FoM: 56.83% vs 51.33%), and Kappa
    coefficient (0.67 vs 0.542). However, the magnitude of improvement should be
    studied alongside the study’s single-city scope discussed in Section 5.4.

This improvement aligns with the findings of previous studies that employed
hybrid RF-based models. For instance, Kamusoko and Gamba (2015) applied
the RF-CA model to binary urban predictions ; and reported a Kappa of 0.51
and FoM of 47%, which outperformed Support Vector Machine -Cellular
Automata (SVM -CA; Kappa 0.39, FoM 39%) and Logistic Regression -
Cellular Automata (LR -CA; Kappa 0.22, FoM 6%), even in simplified two -
class setups.

Further spatial and quantitative comparisons of the 2024 predicted values
highlight these advantages. The actual observed urban area was 505.01km². The
CA-Markov model predicted 457.30km², underestimating the urban area by
approximately 47.71km²; whereas the RF -CA model predicted 441.64km², a
slightly lower estimate, but with superior overall spatial accuracy. Notably, the
RF-CA model correctly identified 343.05km² of urban land compared to
326.43km² identified by the CA-Markov model.

Such deviations suggest that the CA -Markov model may underestimate the
rate of land transformation because it relies primarily on historical transition
probabilities, and does not fully incorporate recent socioeconomic drivers or
dynamic feedback mechanisms into its calculations. In contrast, the RF -CA
model, which integrates updated geospatial and temporal variables, captures
the nonlinear and rapid nature of contempo rary urban growth in Dar es
Salaam more effectively. This underscores the value of hybrid modelling
approaches for rapidly urbanising areas.

Nyatabori Antonietta Kwame & Zakaria Ngereja
20 JGAT Volume 46, Number 2, 2026
5.2 Key Drivers of Urban Sprawl
A feature importance analysis using the RF -CA model provided valuable
insights into the primary drivers of urban expansion in Dar es Salaam.
Population density was the most influential factor, accounting for 23.06% of
the impurity feature importance. This indicates a strong association between
population density and the probability of urban transition. This finding aligns
with urban growth theories , and reflects the city ’s sustained population
growth, which has more than tripled since the late 1980s and had reached
approximately 5.3 million by 2022 (URT, 2022).

Topographic constraints also played a significant role, with an elevation of
13.19%, indicating that urban development was directed toward flatter and more
accessible areas. Additionally, climate variables contributed to the model ’s
prediction of urban transition —with temperature and rainfall contributing
11.75% and 11.45%, respectively—highlighting the influence of environmental
factors on urban planning decisions in tropical regions, such as Dar es Salaam.

Comparisons with previous studies further contextualise this rapid growth.
Bhanjee and Zhang (2018) documented that built-up areas more than doubled
from 255.06km² in 1990 to 532.13km² in 2014 (an expansion of 108.6%).
Similarly, the present study’s observed urban area of approximately 505.01km²
by 2024 compared with the earlier projections by Simon, Lyimo and Yamungu
(2023), who forecasted a built -up area of only 530km² by 2050 using a CA -
Markov model calibrated from 1995 to 2022. These findings emphasise the
accelerated pace of urban expansion in Dar es Salaam beyond historical trends,
reinforcing the need for models that incorporate dynamic drivers.

5.3 Implications for Urban Planning and Policy
The rapid and sustained urban growth revealed by the RF -CA model has
significant implications for urban planning and policy in Dar es Salaam. The
identification of key drivers—particularly population density, topography, and
climate—suggests that planning strategies should prioritise managing
demographic pressures, protecting vulnerable low -elevation areas, and
incorporating climate-resilient designs.

The superior performance of the RF-CA model and its ability to capture nonlinear
growth patterns highlight the importance of adopting hybrid machine-learning-
integrated approaches, and regularly updating datasets. This is particularly
crucial in fast-growing cities where traditional models, such as the CA-Markov
model, may fail to account for recent acceleration of sprawls. Proactive policies
informed by advanced simulations can support sustainable land use allocation,
infrastructure development, and environmental conservation to mitigate
uncontrolled expansion.

CA-Markov and Random Forest-Cellular Automata Models for Predicting Urban Sprawl

21 JGAT Volume 46, Number 2, 2026
5.4 Limitations and Future Research Directions
The study that generated the data for this paper relied on standard accuracy
metrics (overall accuracy, Kappa, FoM, and F1-score) for model evaluation and
30-m Landsat imagery, which limited the detection of fine-scale changes, such
as informal settlements. Resampling population and climate datasets may
introduce uncertainty, and the CA -Markov model ’s assumption of temporal
stationarity is problematic in rapidly changing urban contexts. Socioeconomic
and policy -driven factors have not been fully incorporated into this model.
Accuracy assessment also relied on a random pixel-based training/testing split,
which does not guarantee spatial independence between samples, given the
spatial autocorrelation among neighbouring pixels, potentially leading to
modest overestimation of the reported accuracy, Kappa, and F1 score values.
Additionally, feature importance rankings from the RF -CA model reflect
impurity-based rankings from a single fitted model, and their stability across
random-forest seeds and sensitivity to correlations among predictors were not
evaluated; hence, a permutation-based analysis would strengthen future
interpretation of driver effects.

Moreover, future research could strengthen validation by employing per-class
confusion matrices, spatial error maps , and null model comparisons to
identify prediction errors better. Incorporating dynamic socioeconomic
variables, extending projections to longer horizons, and exploring ensemble
modelling techniques would further refine forecasts in similar rapidly
urbanising settings. Additionally, detailed investigations —using advanced
spatial analytics, landscape metrics, and multitemporal data —into specific
sprawl morphologies —such as leapfrog development, edge expansion, and
infill patterns —could elucidate the drivers of discontinuous growth and
governance influences on urban form. Also, integrating quantitative analyses
with qualitative stakeholder insights and machine learning for automated
pattern classification provides targeted strategies to complement broader
modelling approaches.

  1. Conclusion
    This study provided a comprehensive comparison of the CA -Markov and RF -
    CA models for predicting urban sprawl in Dar es Salaam, Tanzania, using
    multitemporal Landsat imagery and key biophysical and socioeconomic drivers
    from 1995 to 2024. The RF-CA model results demonstrate a better predictive
    performance, which achieved higher predictive accuracy (77.84% vs 69.38%),
    Kappa coefficient (0.67 vs 0.542), and Figure of Merit (56.83% vs 51.33%) than
    the traditional CA-Markov approach. The RF-CA model captured the nonlinear
    spatial complexities of rapid urban expansion more effectively, particularly in
    peri-urban areas ; and identified population density, elevation, and climatic
    factors as the primary drivers of growth.

Nyatabori Antonietta Kwame & Zakaria Ngereja
22 JGAT Volume 46, Number 2, 2026
These findings underscore the value of integrating machine -learning with
cellular automata frameworks for urban growth modelling in data -scarce and
rapidly urbanising African cities, where traditional stationary models are often
inadequate. For urban planners and policymakers in Dar es Salaam, adopting
RF-CA can enable more reliable forecasts to guide proactive infrastructure
planning, zoning regulations, and targeted interventions in high-growth
zones—such as Ubungo, Kimara, and Mbezi —thereby curbing uncontrolled
sprawl and informal settlements. Environmental agencies can leverage these
predictions to monitor and mitigate ecological impacts, including vegetation
loss and wetland encroachment.

More broadly, this study highlights the potential of hybrid geospatial-machine-
learning approaches to support sustainable urban development amid Africa ’s
unprecedented urbanisation, thereby offering a methodological template that
can be tested and adapted for other fast-growing SSA cities facing similar data
and planning constraints.

References
Abatzoglou, J. T., Dobrowski, S. Z., Parks, S. A. & Hegewisch, K. C. (2018).
TerraClimate, A High -resolution Global Dataset of Monthly Climate and Climatic
Water Balance from 1958–2015. Scientific Data, 5(1): 1–12. https://doi.org/10.1038/
sdata.2017.191.
Asif, M. et al. (2023). Modelling of Land Use and Land Cover Changes and Prediction
Using CA-Markov and Random Forest . Geocarto International , 38(1). https://doi.
org/10.1080/10106049.2023.2210532.
Atef, I., Ahmed, W . & Abdel-Maguid, R.H. (2024) . Future Land Use Land Cover
Changes in El-Fayoum Governorate: A Simulation Study Using Satellite Data and
CA-Markov Model. Stochastic Environmental Research and Risk Assessment, 38(2).
https://doi.org/10.1007/s00477-023-02592-0.
Augustijn-Beckers, E.W., Flacke, J . & Retsios, B. (2011) . Simulating Informal
Settlement Growth in Dar es Salaam, Tanzania: An Agent-based Housing Model.
Computers, Environment and Urban Systems , 35(2). https://doi.org/10.1016/j.
compenvurbsys.2011.01.001.
Azabdaftari, A . & Sunar, F. (2024) . Predicting Urban Tomorrow : CA -Markov and
District Evolution. Earth Science Informatics, 17(4). https://doi.org/10.1007/s12145-
024-01340-4.
Batty, M., Xie, Y. & Sun, Z. (1999). The Dynamics of Urban Sprawl. Centre for Advanced
Spatial Analysis, 15.

CA-Markov and Random Forest-Cellular Automata Models for Predicting Urban Sprawl

23 JGAT Volume 46, Number 2, 2026
Behnisch, M., Krüger, T. & Jaeger, J.A.G. (2022). Rapid Rise in Urban Sprawl: Global
Hotspots and Trends Since 1990. PLOS Sustainability and Transformation , 1(11).
https://doi.org/10.1371/journal.pstr.0000034.
Bhanjee, S. & Zhang, C.H. (2018). Mapping Latest Patterns of Urban Sprawl in Dar es
Salaam, Tanzania . Papers in Applied Geography , 4(3). https://doi.org/10.1080/
23754931.2018.1471413.
Breiman, L. (2001). Random Forests. Machine Learning, 45(1). https://doi.org/10.1023/
A:1010933404324.
Fitawok, M.B. & Minale, A.S. (2024). A Review of the Application and Implications of
Cellular Automata -based Urban Growth Model in Africa . South African
Geographical Journal, 106(3). https://doi.org/10.1080/03736245.2024.2308909.
Funk, C., Peterson, P., Landsfeld, M., Pedreros, D. H., Verdin, J. P., Shukla, S., Husak,
G. J., Rowland, J., Harrison, L., Hoell, A. & Michaelsen, J. (2015). The Climate
Hazards Infrared Precipitation with Stations - A New Environmental Record for
Monitoring Extremes. Scientific Data, 2. https://doi.org/10.1038/sdata.2015.66.
Gharaibeh, A. A., Al-Shboul, A. A., Al-Kofahi, S. & Al-Zu'bi, M. (2020). Improving Land-
use Change by Integrating ANN with Cellular Automata -Markov Chain Model.
Heliyon, 6(9). https://doi.org/10.1016/j.heliyon.2020.e05092.
Hamad, R., Balzter, H . & Kolo, K. (2018) . Predicting Land Use/Land Cover Changes
Using a CA -Markov Model Under Two Different Scenarios . Sustainability
(Switzerland), 10(10). https://doi.org/10.3390/su10103421.
Jayabaskaran, M. & Das, B. (2023) . Land Use Land Cover (LULC) Dynamics by CA -
ANN and CA -Markov Model Approaches: A Case Study of Ranipet Town, India .
Nature Environment and Pollution Technology , 22(3). https://doi.org/10.46488/
NEPT.2023.v22i03.013.
Kamusoko, C. & Gamba, J. (2015). Simulating Urban Growth Using a Random Forest-
cellular Automaton (RF-CA) Model. ISPRS International Journal of Geo -
Information, 4(2). https://doi.org/10.3390/ijgi4020447.
Khawaldah, H.A., Farhan, I . & Alzboun, N.M. (2020) . Simulation and Prediction of
Land Use and Land Cover Change Using GIS, Remote Sensing and CA -Markov
Model. Global Journal of Environmental Science and Management, 6(2). https://doi.
org/10.22034/gjesm.2020.02.07.
Laros, M., & Jones, F. (Eds.). (2014). The State of African Cities 2014: Re -imagining
Sustainable Urban Transitions. UN-Habitat.
Li, M., Cao, Y., Dai, J., Song, J., & Liang, M. (2025). A Comprehensive Review of Urban
Expansion and its Driving Factors. Land, 14(8). https://doi.org/10.3390/
land14081534.
Mberu, B., Béguy, D . & Ezeh, A.C. (2017) . Internal Migration and Slums in sub -
Saharan Africa . Africa’s Population: In Search of a Demographic Dividend .
https://doi.org/10.1007/978-3-319-46889-1_20.

Nyatabori Antonietta Kwame & Zakaria Ngereja
24 JGAT Volume 46, Number 2, 2026
Mnyali, E.T. & Materu, S.F. (2021). Analysis of the Current and Future Land Use/Land
Cover Changes in Peri-Urban Areas of Dar es Salaam City, Tanzania using Remote
Sensing and GIS Techniques . Tanzania Journal of Science , 47(5).
https://doi.org/10.4314/tjs.v47i5.12.
Musa, S.I., Hashim, M . & Reba, M.N.M. (2017) . A Review of Geospatial-based Urban
Growth Models and Modelling Initiatives. Geocarto International . https://doi.org/
10.1080/10106049.2016.1213891.
Nugroho, F. & Al-Sanjary, O.I. (2018). A Review of Simulation Urban Growth Model .
International Journal of Engineering and Technology (UAE), 7(4). https://doi.org/
10.14419/ijet.v7i4.11.20681.
Pontius, R.G. & Millones, M. (2011). Death to Kappa: Birth of Quantity Disagreement
and Allocation Disagreement for Accuracy Assessment . International Journal of
Remote Sensing. https://doi.org/10.1080/01431161.2011.552923.
Sang, L. et al. (2011). Simulation of Land Use Spatial Pattern of Towns and Villages
Based on CA -Markov Model. Mathematical and Computer Modelling , 54(3 –4).
https://doi. org/10.1016/j.mcm.2010.11.019.
Schiavina, M., Freire, S., Carioli, A. & MacManus, K. (2023). GHS-POP R2023A: GHS
Population Grid Multitemporal (1975 –2030) [Data set]. European Commission,
Joint Research Centre. doi.org.
Seto, K.C., Güneralp, B . & Hutyra, L.R. (2012). Global Forecasts of Urban Expansion
to 2030 and Direct Impacts On biodiversity and Carbon Pools . Proceedings of the
National Academy of Sciences of the United States of America , 109(40).
https://doi.org/10.1073/pnas.1211658109.
Shaamala, A., Yigitcanlar, T., Nili, A. & Nyandega, D. (2025). Machine Learning
Applications for Urban Geospatial Analysis: A Review of Urban and Environmental
Studies. Cities, 165. https://doi.org/10.1016/j.cities.2025.106139.
Simon, O., Lyimo, J . & Yamungu, N. (2023) . Land Use and Cover Change in Dar es
Salaam Metropolitan City : Satellite Data and CA -Markov Chain An alysis.
GeoJournal, 88(6). https://doi.org/10.1007/s10708-023-10960-0.
Spencer, D. (2009) . Cities and Complexity: Understanding Cities with Cellular
Automata, Agent-Based Models, and Fractals . The Journal of Architecture , 14(3).
https://doi.org/10.1080/13602360903028044.
Tobler, Waldo. R. (1970). A Computer Movie Simulating Urban Growth in the Detroit
Region. Economic Geography (Supplement: Proceedings, International Geographical
Union, Commission on Quantitative Methods. Economic Geography 11(46).
Tumwesigye, S., Vanmaercke, M., Hemerijckx, L.-M., Opio, A., Poesen, J., Twongyirwe,
R. & Van Rompaey, A. (2023). Spatial Patterns of Urbanisation in Sub-Saharan
Africa: A Case Study of Uganda . Development Southern Africa , 40(1) : 1 –21.
https://doi.org/10. 1080/0376835X.2021.1932426.
United Republic of Tanzania (URT). (2022). The 2022 Population and Housing Census:
Administrative Units Population Distribution Report , National Population and
House Census of Tanzania. National Bureau of Statistics, Dar es Salaam, Tanzania.

CA-Markov and Random Forest-Cellular Automata Models for Predicting Urban Sprawl

25 JGAT Volume 46, Number 2, 2026
Wang, S.W., Munkhnasan, L . & Lee, W.K. (2021). Land Use and Land Cover Change
Detection and Prediction in Bhutan’s High-altitude City of Thimphu, Using Cellular
Automata and Markov Chain. Environmental Challenges , 2. https://doi.org/10.
1016/j.envc.2020.100017.
Wang, X., Shi, R . & Zhou, Y. (2020) . Dynamics of Urban Sprawl and Sustainable
Development in China . Socio-Economic Planning Sciences , 70. https://doi.org/10.
1016/j.seps.2019.100736.
White, R. & Engelen, G. (1993). Cellular Automata and Fractal Urban Form: A Cellular
Modelling Approach to the Evolution of Urban Land-use Patterns. Environment &
Planning A, 25(8). https://doi.org/10.1068/a251175.

Nyatabori Antonietta Kwame & Zakaria Ngereja
26 JGAT Volume 46, Number 2, 2026

Appendix

Confusion Matrices
The Appendix reports the year-specific error (confusion) matrices underlying the
Random Forest land cover classification reported in Section 3.2. A set of reference
points was interpreted from high-resolution basemap imagery and compared with
the classified map. Producer’s accuracy (omission error) is read along each row, and
user’s accuracy (commission error) is read along each column. The overall accuracy
and kappa coefficient computed from each matrix are given below the
corresponding table.

Table A1: 1995 RF Land-Cover Classification
Reference \
Predicted
Water Vegetation Built-up Bare
land
Row
Total
Producer’s
Accuracy
Water 9 1 0 0 10 90.0%
Vegetation 1 40 2 3 46 87.0%
Built-up 2 4 36 2 44 81.8%
Bare land 0 1 1 18 20 90.0%
Column Total 12 46 39 23 120
User’s Accuracy 75.0% 87.0% 92.3% 78.3%

Overall Accuracy = 85.83%
Kappa = 0.796

Table A2: 2004 Rf Land-Cover Classification
Reference\
Predicted
Water Vegetation Built-up Bare
land
Row
Total
Producer’s
Accuracy
Water 11 3 0 0 14 78.6%
Vegetation 3 39 3 2 47 83.0%
Built-up 0 1 35 1 37 94.6%
Bare land 0 5 1 16 22 72.7%
Column Total 14 48 39 19 120
User’s Accuracy 78.6% 81.3% 89.7% 84.2%

Overall Accuracy = 84.17%
Kappa = 0.774

CA-Markov and Random Forest-Cellular Automata Models for Predicting Urban Sprawl

27 JGAT Volume 46, Number 2, 2026
Table A3: 2014 Rf Land-Cover Classification
Reference \
Predicted
Water Vegetation Built-up Bare
land
Row
Total
Producer’s
Accuracy
Water 13 0 0 0 13 100.0%
Vegetation 4 33 4 3 44 75.0%
Built-up 3 5 37 1 46 80.4%
Bare land 0 0 0 17 17 100.0%
Column Total 20 38 41 21 120
User’s Accuracy 65.0% 86.8% 90.2% 81.0%

Overall Accuracy = 83.33%
Kappa = 0.76

Table A4: 2024 Rf Land-Cover Classification
Reference \
Predicted
Water Vegetation Built-up Bare
land
Row
Total
Producer’s
Accuracy
Water 14 1 0 0 15 93.3%
Vegetation 0 39 3 1 43 90.7%
Built-up 2 4 34 4 44 77.3%
Bare land 0 0 0 18 18 100.0%
Column Total 16 44 37 23 120

User’s Accuracy 87.5% 88.6% 91.9% 78.3%

Overall Accuracy = 87.50%
Kappa = 0.824

论文原文与出版信息

Journal of the Geographical Association of Tanzania
查看论文原文

继续探索