The study area (Figure 1) falls in the Moist Semi-deciduous Northwest Forest type zone and includes the old Kumasi Metropolitan area and Atwima Nwabiagya District in the Ashanti Region. It has the Owabi-Barekese headworks. The Owabi-Barekese watershed is the main source of treated water for the Kumasi Metropolis and its surrounding communities. The Owabi (wildlife sanctuary

and a Ramsar Site) and Barekese catchment were designated in 1920 and 1972 correspondingly through an Executive Instrument (Forestry Commission, 2014; Akoto et al., 2021) . A forest cover within the catchment area was created to safeguard recharge areas, prevent siltation and rapid evaporation of water in the reservoir as well as improve rainfall. The catchment area has rich diversity of vegetation and wildlife and supports critical ecosystem services (Forestry Commission, 2014; Amuquandoh et al., 2011; Koranteng, 2017) .

The major software employed in this study includes (Environmental Systems Research Institute (ESRI) ArcGIS 10.0, ERDAS Imagine and IDRISSI Selva. This study is based on the data sources listed in Table 1.

Preprocessing of satellite images is vital and seeks to harmonize data and the biophysical phenomena it signifies (Parsa et al., 2016) . Pre-processing was done utilizing ArcGIS 10.0. All the images were re-sampled to 30 × 30 meter pixel resolution to make accurate analysis of the datasets and comparability possible.

Five LULC categories—Close_forest, Open_forest, Agriculture, Built_up and Water (Table 2) were chosen and used in the image classification (supervised classification) based on the authors’ local knowledge of the study area and literature (Koranteng et al., 2020; Frimpong & Molkenthin, 2021) . Random Forest Algorithm was employed to allocate pixels to their classes since the random forests are not limited by statistical assumptions.

The Random Forest technique uses the construction of trees to classify satellite image data (Breiman, 2001) . The algorithm has a lower computational time and a higher classification accuracy (Rodriguez-Galiano et al., 2012; Inglada et al., 2015) . The spectral bands were the input data, and the LULC class with the most segments was the output. Due to the independence of each tree, training and testing were carried out simultaneously (Figure 2) (Breiman, 2001) . Given a collection of multispectral images (spectral bands), denoted as X1, X2, …, Xn, where n denoted the sample count and L denoted the number of features, While “z” was the link between sample “I” and sample “xj,” where “z” Z = “1, −1,” “Xi” indicated the position of sample I in the space “RL × n.” Z = 1 if xi and xj belonged to the same class.

Tree-structured predictors are present in the random forest. The collaboration formula for the kth tree was f_k(x) = f(x, k), where f(x, k) was a tree that grew in size as the training set and random vector k did, capturing the tree’s many stochastic building blocks. Each tree cast a vote for the most popular LULC class, and the k was independent and evenly distributed (Breiman, 2001) .

For a sample, the probability of correctly predicting class “z” was

P ( z | x ) = 1 K ∑ k = 1 K P k ( z | x ) (1)

where P_k(z|x) was the projected density of class labels of the kth tree and K was the number of trees in the forest. As illustrated the decision function of the forest is given as

C ( x ) = argumentsofthe   maxima j ∈ Z P ( j | x )

The margin function for the RF was

m l ( x , z ) = P ( z | x ) − max j ∈ Z j ≠ z P ( j | x ) (2)

If ml(x, z) > 0, might result the correct classification. The value of generalization error is limited as Random Forest can regulate without increasing the trees

(Amit & Geman, 1997; Ye et al., 2013) . The convergence function at this point can be explained as;

m l ( x , z ) = P ( z | x ) − max j ∈ Z j ≠ z P ( j | x ) < 0 (3)

indicating that the generalization error decreases as the number of trees grows. The generalization error was given an upper bound by

P E ≤ ρ ¯ ( 1 − s 2 ) / s 2 (4)

where s is the strength of the set of f(x, k), PE is the generalization error, and is the average value ( ρ ¯ ) of the tree correlation.

The generalization error of the RF converged to a fixed value once there were enough trees in the forest. Each tree produced a “yes” or “no” vote from the RF (Figure 2).

Accuracy assessment was performed on the 2020 satellite image, an assessment report was produced as Confusion Matrix (Table 3). For 1990, 2000 and 2010, these classified images were validated using ancillary data (Table 1).

LULCC post-classification detection method adopted using the ERDAS Imagine, which entailed using two classified images to make a comparison to produce change information. Consequently, the differences between two images represent the change. The extent of change and percentage of changes can be stated in a straightforward formulation as used by Mahmud & Achide (2012) and Hua (2017) :

K = F − I (5)

A = ( F − I ) x 2 2 ! I × 100 (6)

where K is extent of changes, A represent the percentage of changes, F is first data, and I is reference data.

Prognostication of LULC changes for 2040 used IDRISI Selva. This research utilized LULC methods in RS to ascertain differences and explain the percentage of LULC changes in that time, along with the approximation for the subsequent 30 years.

Markov chains are stochastic processes (Caraka et al., 2019; Nemeth & Fearnhead, 2021; Shepero & Munkhammar, 2018) and the matrices indicate variations between land use types (premised on the assumption of perpetuation of past development) (Ghalehteimouri et al., 2022; Rahnama, 2021) and are often employed in simulation variations and developments of LULC (Chang et al., 2021; Sobhani et al., 2021) . The consistency of Markov model for projection of LULC alterations can be accurately described in this mathematical way (Subedi et al., 2013; Okwuashi & Ndehedehe, 2021; Das & Sarkar, 2019) :

L ( t + 1 ) = P i j × L ( t ) (7)

P i j = [ P 11 P 12 P 1 m P 21 P 22 P 2 m P m 1 P m 2 P m m ] (8)

where L_(t) and L_(t+1) represent land use status at time t and t + 1 respectively. Including ∑ j = 1 m P i j = 1 (i, j = 1, 1, 2, m) is the transition probability matric in a given state.

In this paper, 2010 and 2020 maps were used in Markov chain model to create the transition matrix changes between the existing 10 years, and the process is repeated onto map 2040 for upcoming land use to obtain the transition matrix changes in IDRISI Selva v.17 environment.

Ulam and Von Neumann in the 1940s originally proposed Cellular Automata (CA) for usage in land use changes theoretical study (Nugraheni & Natali, 2018; Xing et al., 2020) . CA fundamental theory is summarized as the land use modifications for any place (cells) can be clarified by the present state and variations in neighboring cells (Mansour et al., 2020; Noszczyk, 2019) .

Markov Chain and CA models are reckoned as valuable for predicting land use changes (de Oliveira Barros et al., 2018; Munthali et al., 2020) . The Markov-CA model employs Markov Chain outputs to utilize a contiguity filter to facilitate the advancement of other land use types from one time to another (Parsa et al., 2016) . CA then advances a not random spatial weighting on the places which have approximately the same features to the existing land use category based on classes (Subedi et al., 2013) . Consequently, Markov-CA model is deemed a vigorous tool as it provides temporal, spatial and quantitative estimation and LULC dynamics in modeling for the future (Mishra & Rai, 2016; Subedi et al., 2013) .

Figure 3 summaries the flowchart of the study.

LULC maps generated for the years1990, 2000, 2010 and 2020 from supervised classification which used the categories Closed Forest, Open Forest, Agriculture, Built-up and Water are shown in Figure 4. Foody (2002) ; Behera et al. (2012) theorize that accuracy assessment is vital when post classification change detection method is employed. The accuracy assessment conducted on the LULC Map 2020 is shown in Table 3. 98% represented the total classification accuracy and 0.97 achieved as the overall Kappa statistics. For the epoch years 1990, 2000, 2010, topographical map, aerial photograph, land cover map, digitized topographical data photographs, data and reference points from statutory bodies such as the Forestry Commission and Survey Department of the Republic of Ghana were used to check for accuracy assessment averaging 90%.

In 1990 map, the entire area was covered by forest predominantly, open forest (58.31%) with close forest (7.85%). Aside from forest cover, agricultural lands

occupied 22.52% with the urban/built-up areas covering 11.15% whilst the waterbodies occupied a marginal 0.17% of the total area of 81838.9 ha. In the year 2000, both the close and open forest experienced reductions whilst agricultural lands increased. However, a substantial increase in built-up areas was observed. Waterbodies category experienced increases from 0.17% (1990) to 0.49% (2000). There was a rise in both the close and open forests in 2010. Conversely, agricultural lands decreased significantly but built-up areas also increased. The waterbodies again increased marginally in 2010. In the last year of study (2020), there was an increase in the close forest contrary to open forests that decreased. However, the dwindling of agricultural lands continued. Nonetheless, there was a substantial rise in the built-up areas (Table 4 and Figure 4).

The close forest decreased from 1990 to 2000 but increased from 2000 to 2020. Averagely within the study period, the close forest decreased by 1.95%. The open forest decreased from 1990 to 2000 but there was an increase from 2000 to 2010. There was once again a decrease from 2010 to 2020. Generally, there was a decrease of −20.82% from 1990 to 2020. There was an increase from 1990 to 2000 agriculture class. On the other hand, a reduction in agricultural lands was observed from 2000 to 2020. Largely there was a decrease from 1990 to 2020 (−5.14%). The built-up areas intensified throughout the study period. In the case of the waterbodies, there was an increase from 1990 to 2010. It decreased from 2010 to 2020. Furthermore, there was an overall increase of 0.27% in the waterbodies (Table 5 and Figure 5 and Figure 6).

The simulated 2020 LULC map showed slight variation in the distribution and values compared to the actual 2020 LULC map (Table 6). This result indicated a very robust method for LULC map stimulation as there were marginal differences. The Statistical evaluation based on the Kappa coefficient was used to

measure the overall agreement of the matrix, the ratio diagonal values’ summation versus the total number of pixel counts within the matrix and the non-diagonal elements would be the best approach to assess the model accuracy (Arsanjani et al., 2013) . The accuracy assessment obtained K values (Kstandard = 0.7595; K no = 0.8913; K location = 0.8841; KlocationStrata = 0.8241). These values agree with Monserud & Leamans (1992) standard values stipulate that a kappa value of 75% or greater indicates excellent classifier performance, as 40% value or less than is poor.

The simulated LULC map for 2040 is shown in Table 7 and Figure 7. The cumulative forest category (Close and Open forests) had dwindled as the area has largely been transformed into built-up. Agriculture class decreased as transformed into built-up. Built-up areas and water increased to 45.15% and 0.61% respectively. Built-up category assumed the highest LULC class of the total area in 2040.

The RFC utilized for the classification of satellite images and the efficiency of this algorithm as used in the works of (Frimpong & Molkenthin, 2021; Nyamekye et al., 2021) was superior, when contrasted with the traditional classifiers used in the classification satellite images. The kappa, producer and user accuracies yielded satisfactory results and agreed with works by Abass et al. (2018) ; Buo et al. (2021) .

The LULC maps showed that most built-up areas were in the south-eastern parts of the study areas in the Kumasi metropolis, which is almost completely denuded of natural vegetation. The diminishing agricultural lands could be ascribed to the heightened demand for lands for commercial centres and residential areas. The analysis of LULC revealed that forestlands have been subjected to intense pressure due to anthropogenic activities (Frimpong, 2015) . The conversion of forestlands and agricultural lands into urban/built-up areas is considered rewarding in the Ghanaian setting (Abass et al., 2018; Buo et al., 2021; Wemegah et al., 2020) . The rapid rise in the population is a perculiar reason that could have caused the growth in the built-up areas, and it endorsed by the outcome of the study of (Ghana Statistical Service (GSS), 2013; Addae & Oppelt, 2019) . The results support the findings in previous literature that the research utilized the same LULC classes in earlier works (Frimpong & Molkenthin, 2021; Koranteng et al., 2015) . The populace living in the peri-urban areas tends to rely on forest resources for sustenance (Cobbinah et al., 2020) . It was found out that forestlands are the foremost to be transformed to built-up areas. This outcome is supported by the work of (Frimpong, 2015) . Consequently, urban spatial growth is largely attributed to the comparatively lower prices of land at the urban fringes for residential purposes (Wellington, 2009) . This could be the reason for the urban/built-up areas growing outwards from the already existing urban/built-up centers.

The forest cover (especially the Open Forest category) was the most depleted LULC category under the study period. Forest loss has always been a crucial problem in Ghana even before independence due of the different forces behind it (Benneh & Agyepong, 1990) . Forest lands forcefully appropriated from indigenous landowners and families for various reasons by the colonial master’s forest policies led to the wanton exploiting the forest cover haphazardly (Agbosu, 1983) . After independence (1960-1970), deforestation was intensified due to the cultivation of cocoa as an export product (Dei, 1992) . From 1981 to 1985 timber had become the primary driver of deforestation as it was the third largest export article of trade providing 5% - 7% of Gross Domestic Product (GDP), (International Institute for Environment and Development (IIED), 1987; Owusu, 2010) . The period 2003-2013 has been characterized by plantation establishments which resulted in incremental surge in the annual rate of forestlands loss (Davis & Philips, 2005; Afele et al., 2022) . The period 2013-2022 has experience intensification of forest loss due to unbridled urbanization (Kyere-Boateng & Marek, 2021; Ofori et al., 2022) .

Anthropogenic activities and actions are the immediate causes of loss (deforestation and forest degradation). The underlying forces and causes of forest loss are multi-faceted, complex and differ from place to place. The drivers of the rapid LULC in the study are largely attributed to unbridled urbanization, Population growth and Settlement, Infrastructural Development, Mining and Unsustainable Agricultural Practices.

The Markov-CA model employed in this research was observed to be satisfactory and agrees with other studies (Hasan et al., 2020; Fitawok et al., 2020; Kushwaha et al., 2021) . The Kappa value of 85% was achieved in accuracy assessment. The result is supported by several papers and is indicative of a robust model (Rwanga & Ndambuki, 2017; Mondal et al., 2016; Kaul & Sopan, 2012) . However, the prognosis for 2040 is not encouraging for forest cover protection. This result is indicative of an unbridled urbanization and agrees with earlier study by Koranteng & Zawila-Niedzwiecki (2015) in the Ashanti Region of Ghana where Built-up class surge was very intense and buttressed by studies from Cobbinah et al. (2015) and Obeng-Odoom (2014) which states that Ghana’s population tend to amass in urban areas.