2.1 Background of study area

Kuantan is located in the eastern part of Peninsular Malaysia at the latitude of 3°48′27.72″N and longitude of 103°19′33.60″E and is the capital city of Pahang with a population approaching half a million people (DOSM, 2021). The city is rich with local culture, known for its natural and heritage tourism attractions, and also has abundant natural resources. Kuantan's economy was originally dependent on agricultural activities (comprising oil palm plantations, rubber plantations and orchards), marine and freshwater fisheries, small businesses such as craft production and small-scale manufacturing, and also local tourism activities (ECER, 2016). In recent decades the economy has been industrialising and the five key future economic drivers focus of the ECER master plan are: (a) tourism, (b) oil, gas and petrochemical, (c) manufacturing, (d) agriculture and agro-based industry and (e) human capital development (ECER Master Plan 2.0, 2019; ECERDC, 2010). Additionally, the recent discovery of Kuantan land being rich in gibbsite minerals has also led to the acceleration of mining activities (Head, 2016). The study area (Figure 1) includes key drivers of the rapid development within the region: the growing industrial sectors, vast agricultural activities, bauxite mining excavation and also the future ECRL. It also includes some forest reserves to sustain the local biodiversity and protected species (Chyuan, 2021). The total study area is 698.9 km², with a length of 29.6 km and a width of 27.3 km. The boundary was chosen as it included important and relevant land-use change considerations for the region.

2.2 Methods overview

The analysis utilises remote sensing data to produce a historic series of LULC maps for the years 2010, 2015 and 2020 through image preprocessing and Random Forest classification using Google Earth Engine (GEE) and ArcGIS. The classified LULC maps were refined further with finer land-use categories using various sources of auxiliary reference data (Lechner, Owen, et al., 2019). Land-use change analysis and future land-use modelling were conducted using the LCM modelling system in the TerrSet software. Finally, an accuracy assessment was carried out to determine the reliability of the classified historical LULC maps and the simulated future LULC maps.

2.3 Remote sensing of historic LULC

The satellite imagery for this study was derived and analysed on the GEE platform (Sidhu et al., 2018). Its advanced analyses and cloud computing capabilities are particularly well-suited to perform multiple map analysis at greater speeds and capacity (Gorelick et al., 2017; Mutanga & Kumar, 2019). The satellite imagery from Landsat 8 OLI (for 2015 and 2020) and Landsat 5 TM sensors (for 2010) (Supporting Information Appendix A.1) were preprocessed to create a multitemporal composites of dates that provided fair coverage of key periods of land-use change for the different sectors (i.e., plantation, industrial development and bauxite mining) in the study area. The choice of years was influenced by the availability of the least cloudy scenes (with a maximum of 5% cloud cover). The multitemporal composite was composed of the median pixel value from atmospherically corrected and cloud masked Landsat satellite surface reflectance data for the chosen 2010, 2015 and 2020 LULC maps (Ang et al., 2021) (Supporting Information Appendix A.2).

A multistep land-cover classification was carried out to characterise the LULC classes related to key existing and emerging land-use development activities based on two major steps: (1) a supervised classification to identify the coarse-scale land covers and (2) manual digitisation using auxiliary reference data to facilitate the identification of fine-scale LULC which are significant drivers of land-cover change in this study (Figure 2). The first step involves the categorisation of coarse-scale land covers that could be distinguished through supervised classification, such as bareland, waterbody, built-up, mining sites, forest and secondary vegetation. Carrying out the supervised classification with Landsat data had some limitations, as it could not precisely distinguish fine-scale classes that share similar spectral properties. For example, distinguishing different vegetation types, such as oil palm agricultural land from other disturbed vegetation, or differentiating among different built-up types (i.e., residential areas with industrial sites). Hence, a second step was required, where some of the broad-scale classes were divided into fine-scale classes through manual attribution with the aid of auxiliary data (Supporting Information Appendix A.2).

Random Forest machine learning algorithm was used in GEE to carry out the supervised classification of the preprocessed multitemporal composite imagery to distinguish the coarse-scale LULC classes that frame the dominant land covers (Table 1). These coarse-scale classes were identified through the direct observation of the physical material at the surface of the earth (Fisher & Unwin, 2005), covering various homogeneous spectral land-cover types.

Auxiliary data (Supporting Information Appendix A.2) were used to derive training and accuracy assessment points; a commonly practiced method for accuracy assessment (Foody, 2002; Liu et al., 2002; Plourde & Congalton, 2003). 0.5–30-m resolution true-colour data in Google Earth Pro was predominantly used for ground truthing, while the assessment of primary forest cover was supported by the Global Forest Change 2000–2014 data from Hansen et al. (2013). Additionally, the ArcGIS base map with a higher resolution at 0.5 m was also used specifically for the classification of the 2020 map. A key challenge for classifying historical data (i.e., 2010 and 2015) is the unavailability of high-resolution true-colour imagery to adequately identify the land covers, hence, ground truthing using other auxiliary data was necessary. To interpret these data, a range of false-colour band combinations (Supporting Information Appendix B) was used to clearly distinguish between different land covers (Quinn, 2001). Using these data sets, a minimum of at least 130 training and accuracy points for each of the six coarse-scale land-cover classes were identified.

The training and accuracy points assigned for the coarse-scale classes were divided using a ratio of 70:30 where 70% of the validation points were assigned for training the Random Forest classification, while the remaining 30% were reserved for accuracy assessment (Supporting Information Appendix C). There were a total of 546 training points and 234 accuracy assessment points. The output generated for accuracy assessment was interpreted as overall accuracy, producer's accuracy and user's accuracy. The produced classified imagery was then refined with the aid of auxiliary reference data to produce fine-scale LULC map (Supporting Information Appendix A.2).

The second step involves a post-classification process (Figure 2), where we manually refined and updated the produced LULC classification map with fine-scaled classes, as shown in Table 1. Using additional ancillary data to update existing classes (i.e., using linework from another data set) and using auxiliary data to support manual digitising (Supporting Information Appendix A.2) the finer-scale classes were mapped. We used the annual oil palm plantation maps from 2001 to 2016 by Xu et al. (2020) as a reference to digitise the oil palm plantation extent with the addition of true-colour imagery. To improve the accuracy of roads, vectors from Open Street Map were used. Starting with the 2020 LULC map, these two classification steps (Figure 2) were repeated, by consecutively taking into account the previously classified LULC maps. This procedure was essential to ensure consistency in the LULC classification process (Ang et al., 2021). As a final product, a series of LULC classification maps were produced for the years 2010, 2015 and 2020.

2.4 Land-use change modelling

In this study, we used one of TerrSet software tools called LCM for analysing the historical land-cover change and MLP neural network and Markov chain analysis modelling to model the land-use change transition probabilities and produce future LULC change projections (TerrSet Geospatial Monitoring and Modelling Software, 2021) with the following three steps: (1) change analysis, (2) model validation and (3) change prediction as shown in Figure 3.

First, the historical LULC maps were analysed as pairs (i.e., 2010 with 2015; 2015 with 2020) to obtain a transition area matrix to highlight the dominant land-use transition between each class in terms of area (Megahed et al., 2015). Next, the LULC transition potential maps produced from step 1 are used in step 2 to produce the transitional probability scenarios that estimate the probability of each pixel's persistence or conversion to another land cover for further land-use change modelling (Dzieszko, 2014), with the assistance of MLP neural network trained by a range of explanatory variables (Figure 4) (Eastman, 2006). These explanatory variables were selected based on their potential influence on the development activities, such as industrial site expansion, mining and urban sprawl (Figure 4 and Supporting Information Appendix D).

The explanatory variables mainly consisted of topographic and proximity factors. Topographic factors such as the digital elevation model, slope and landform data set acquired from previous studies (Farr et al., 2007; Theobald et al., 2015) are generally considered one of the most significant factors affecting urban sprawl, influencing city size and spatial distribution (Hasan et al., 2020). The population density data set from Gaughan et al. (2013) was also included to account to model urbanisation growth. Additional explanatory variables including soil and lithology data sets were sourced from the studies of Law (1968) and Hartmann and Moosdorf (2012). The MLP neural network trained with these explanatory variables was then used to assess the relative power of the explanatory variables through LCM's backward stepwise analysis by consecutively eliminating the weakest explanatory variable to find the best combinations of explanatory variables that influence the transition probability of each LULC class.

Following step 2, the previously produced transition probability scenarios are treated as a foundation from which the LCM would model the future LULC projection through the Markov chain analysis to estimate the expected quantity of change using a competitive land allocation model to determine the potential future scenario. Once the future LULC projection is made, validation can be carried out, whereby the quality of the future LULC modelling is assessed by comparing the simulated future LULC map against the actual LULC map. Simply stated, the 2020 LULC projection was simulated using the 2010 and 2015 LULC maps and then compared to the actual classified 2020 LULC map. At this stage, two versions of predictions are made by the model: (1) hard prediction and (2) soft prediction. The hard prediction map projects the land-cover map with each class representing their most likely probability of land-cover class in 2020 (Ayele et al., 2019; De Alba & Barros, 2012). The soft prediction projects a vulnerability map, determining the probability of the pixels changing to another land category (Adepoju et al., 2006; Ayele et al., 2019; De Alba & Barros, 2012). Step 3 is a reiteration of step 2, but excludes the validation process; using the years 2015 and 2020 LULC maps to derive a 2030 LULC projection map (Figure 3).

Next, we used multiple metrics to validate the accuracy of the LULC modelling outputs (in step 2) based on differences between the simulated and the actual map of the year 2020. The first was the Kappa Index Agreement (KIA) approach, as an overall accuracy evaluation, to measure the accuracy of the modelling system by accounting for all the transitions and persistence of the pixels during the MLP training process (Eastman & Toledano, 2018). We calculated $$ which indicates the overall agreement of the simulation, $$ which indicates the extent to which the two maps agree in terms of the location of each category, and $$ which indicates the proportion of simulation assigned correctly versus the proportion that is correct by chance (Kitada & Fukuyama, 2012; Mishra & Rai, 2016). Following Pontius and Millones (2011), who suggested that Kappa scores can be misleading, we used two alternative metrics as well. These metrics were (1) Areas Under the Curve (AUC) of Total Operating Characteristic (TOC) and (2) the agreement and disagreement score parameters. Agreement and disagreement were calculated at both grid cell level and due to quantity, and also by the agreement chance (Azari et al., 2022; Leta et al., 2021). The TOC method indicates how well a model predicts change based on the proportion of true event cells in the simulated map, while, the disagreement scores provide proof of the disagreement level between the actual and simulated map (Leta et al., 2021; Shade & Kremer, 2019). Here, the disagreement at the grid cell level refers to the level of the simulated map failing to determine the correct locations of the LULC categories, while, disagreement due to quantity is associated with the level of the simulated map failing to correctly quantify the proportion of each LULC category according to the actual map, and vice versa for the agreement parameters (Shade & Kremer, 2019).

In the final step of the analysis, we overlaid reserve boundaries and calculated how the land cover changed within the reserves for both the historical imagery and modelled future LULC. The reserve boundaries were derived from Global Forest Watch (https://www.globalforestwatch.org/) and Hutan Watch (https://www.hutanwatch.com/) publicly available data on forest reserves and protected areas. Data on the reserves in the study area indicated that they did not include production forestry reserves.

3.1 LULC classification map and accuracy assessment

The classified LULC maps for the years 2010, 2015 and 2020 are shown in Figure 5, while Figure 6 delineates the area (km²) of the LULC classes in their respective years. The spatial distribution of the LULC in the study area changed continuously throughout the years, with the largest LULC extents made up of oil palm plantations, forest cover, disturbed vegetation and waterbodies which includes the ocean and standing freshwater sources. These classes are also the LULC that have experienced the most land-use change activities over time (Figure 6). Towards the southeast region of the study area, minor LULC classes (Figure 5) such as industrial sites, residential and mining sites are clustered together within a highly developing and dynamic region. The accuracy assessment for the classified LULC maps takes into account the coarse-scale land-cover classes, namely, built-up, bareland, waterbody, forest cover and disturbed vegetation. The calculated accuracy score indicates a reasonable level of overall accuracy at an average of 83%. In some cases, the accuracies were as low as 60% for bareland in the year 2020, and 66% for built-up in the year 2010 due to homogeneous spectral properties and human error in classification. (The complete accuracy assessment matrix is presented in Supporting Information Appendix E.)

3.2 LULC change analysis

3.3 Future LULC change projection and validation

Modelling sensitivity was quantified through the backward stepwise analysis that eliminates the weakest explanatory variables one by one to test the influence of the explanatory variables on the skill measure and accuracy of the modelling system to predict class transitions and persistence. The results in Supporting Information Appendix H shows that the model with all 12 explanatory variables is the best combination that resulted in an accuracy rate of 58.19% and 0.5471 of skill measure. Hence, all 12 layers were shown to significantly contribute in training the MLP neural network in producing the transition probability maps.

Model validation was conducted to determine the quality of the 2030 simulated map by first evaluating the difference between the simulated and actual LULC map in the year 2020. The visual comparison between the classified and simulated LULC maps for the year 2020 is shown in Supporting Information Appendix I and the difference in area statistics of all the LULC categories for both maps are presented in Supporting Information Appendix J. On the basis of Supporting Information Appendix J, most of the LULC classes share a similar trend in land-use area, except for the forest cover and oil palm plantation which experienced both underestimation and overestimation, respectively, in the simulated map. The simulated map also poorly estimated the extent of mining sites as per the 2020 actual LULC map and projected more mining sites to grow near the industrial region (Supporting Information Appendix I), although statistically the differences in land-use area between the simulated and classified LULC maps were low (Supporting Information Appendix J).

The KIAs for both the actual and simulated 2020 maps were: $$ = 0.7746, $$ = 0.8135, $$ = 0.7493. The K values scored an average of 0.7791 (77.91%), which is considered reasonable and shows a substantial agreement level between the actual and simulated maps, as the overall K value exceeded 70% (Viera & Garrett, 2005). The other results for the additional modelling validation measuring components are presented in Table 3 for the agreement and disagreement scores and Supporting Information Appendix K. The AUC result of 0.59, suggest a reasonable level of accuracy. The agreement scores (Table 3) also show an overall good agreement between the actual and simulated map (79.71). While the disagreement scores were generally low and this is mainly due to the allocation model of the grid cells (13.90%) rather than the quantity errors (6.39%), coinciding with modelling errors such as the misallocation of mining sites as per the visual interpretation from Supporting Information Appendix I. The results show that the modelling has a good ability to predict LULC changes in terms of overall quantity than the exact location.

The projected future LULC map for the year 2030 is shown in Figure 7 in two forms: hard and soft projection. The hard projection illustrates the potential land-use scenario based on the most likely future land-use change activities, while the soft projection shows a continuous map of the vulnerability of an area to change based on the degree to which a pixel belongs to each of the land-use class (Eastman, 2016). While the statistical analysis of the projected 2030 LULC map in comparison to the earlier classified LULC maps (2010, 2015 and 2020) is shown in Figure 6. Overall, the model predicts an increase among all the LULC classes except for the forest cover and disturbed vegetation which are decreasing; though the rate of decrease is reducing with time. The model simulated a large increase in mining sites and oil palm plantations, by 30.41 and 283.61 km², respectively. These two classes are projected to expand further towards the east of the study area and further encroach on residential areas and industrial sites (Figure 7). The 2030 projections in Figure 7 indicate that the southeast region of the study area will be more vulnerable to land-use change activities; mainly involving the rapid expansion of mining sites, oil palm plantations and reduction in forest cover.

The assessment of forest change within the reserve areas showed a steady decline in forest extent in both the historical and the 2030 land-use change scenarios (Figure 6). At the start of our analysis in 2010, forest cover represented 89% of the reserve extent, which then declined by around 21% in 2020 to 61.24 km², and the future scenario predicts it will decline to 46.74 km².

4.1 Current LULC change analysis

Our analysis showed large spatial and temporal changes throughout the different years of classified LULC maps, especially for the forest cover, oil palm plantations and disturbed vegetation, as these three classes were observed to consistently experience trade-offs between them. Oil palm plantations had the largest extent and the greatest overall change in land use in the study area region (Figure 6), which are in part driven by the nearby palm oil processing and exporting facilities at Kuantan Port (Shevade & Loboda, 2019). Oil palm is recognised as a major economic sector in the ECER region, which contribute to a significant proportion of gross domestic product growth from current and increasing production (ECERDC, 2010). The next largest land cover in the region is forest and disturbed vegetation, which have experienced a consistent reduction through conversion to oil palm, bareland and consequently to other LULC classes. Here, disturbed vegetation and bareland represent transitional land-cover classes (Supporting Information Appendix G). Major deforestation was also observed in the largest patch of forest (Figure 5), creating forest fragmentation that would likely pose a threat to existing wildlife (Chyuan, 2021). The extensive conversion of natural forests to oil palm, such as in this study area, is a key driver of biodiversity loss in the region (Shevade & Loboda, 2019; Wilcove & Koh, 2010). The consistent and significantly diminishing vegetation cover in the study area is likely to persist into the future, based on the current trend shown in the land-use change analysis and future projection. The study also showed a steady decline in forest extent in both the historical and 2030 land-use change scenarios within forest reserves, for an overall historical loss of 37% forest extent (Figure 6).

Kuantan is a dynamically evolving arena where various developments are occurring in parallel, as it falls in a Malaysian development hotspot, one of the Key Development Areas (KDAs), also known as the ECER Special Economic Zone. The Malaysian government's strategy of establishing KDAs focusing on the region's economic strength and rich resources through encouraging industrialisation and commercialisation, and the creation of townships such as by the Federal Land Development Authority (FELDA) (Alden & Awang, 1985; ECER Master Plan 2.0, 2019), can be clearly seen in the past land-use change pattern analysis. The south-eastern part of the study area (Figure 5), near the Kuantan Port, was the most dynamic as residential, mining and industrial sites grew in closer proximity to each other. The strategic location of the Kuantan Port, as a gateway to the Asia-Pacific markets, has brought investment and trade opportunities, including high-profile manufacturing industrial sectors following bilateral trade relations between Malaysia and China (Nur Fatin, 2016; The Report Malaysia, 2012). As a result, over the years of LULC change analysis, land-use changes related to the mega industrial parks and port facilities can be seen. While, a boom in bauxite mining in the study area region from 2015 onwards is part of Malaysia's ambition to be the largest bauxite supplier to China after Indonesia's ore exports ban, where bauxite shipments peaked at nearly 3.5 million tonnes a month (Chow, 2017). These development activities may improve livelihoods by creating more than 200,000 jobs and advancing human capital development (ECER Master Plan 2.0, 2019), driving the expansion of local settlements. The various development agendas can be observed in the past land-use change maps in Figure 5.

The land-use change trends have been driven by the urbanisation and industrialisation activities of the Kuantan region, which if not sustainably managed, may negatively affect ecosystem services and potentially exacerbate pollution while inducing land-use conflicts (Lechner, Verbrugge, et al., 2020; Lourdes et al., 2021). Land-use conflicts occur when there are incompatible interests among stakeholders as a result of development that has conflicting negative effects on the surrounding environment and society (Von Der Dunk et al., 2011). Land-use conflicts in coastal regions like Kuantan are likely to be high, as mega infrastructure projects will compete with coast-dependent economic activities like fisheries, tourism and residential development, as well as environmental, agriculture and mining interests (Susman et al., 2021). Such contrasting land uses are key challenges for land-use planning and decision-making globally (Everingham et al., 2018; Hilson, 2002; Ocelík et al., 2019).

Although the observed land-use expansion for both the industrial and mining sites in this study is not as large as other relatively more extensive land covers such as the oil palm plantation (Figure 6), the impact on the environment from these land uses is much greater. The spill-over environmental effects from industrial and mining locations in the region have been shown to impact negatively on the surrounding natural environment and residential areas, affecting people's health both directly and indirectly (Hossain et al., 2012), the livelihood of fishermen and the tourism sectors (Aw & Awale, 2015; Sobahan et al., 2013). According to a recent study by Hossain et al. (2012), water bodies near the industrial areas in Gebeng, namely, at Pengorak, Tunggak and Balok rivers and Pengorak beach, have had abnormally high levels of heavy metal and aluminium concentrations, with traces of radioactive materials, suspected to originate from the nearby industrial and bauxite mining sites. Other similar studies have found that the rivers located in the vicinity of industrial sites contain higher contaminants, mainly from industrial pollutants, and rivers have been degraded (Ata et al., 2018; Sobahan et al., 2013; Yaakub et al., 2018). Local residents are concerned over impacts from the operation of a rare earth processing plant known as Lynas Advanced Material Plant at Gebeng, as the plant generates waste products that contain low-level radioactive materials, and therefore could impact the health of 700,000 people living within a 30-km radius of the plant and has the potential to contaminate the surrounding natural environment, especially groundwater (Bodetti, 2019; New Straits Times, 2019; Raman & Abdul Kader, 2019; Lipson & Hemingway, 2019).

The consequences of land-use change are complex and interrelated, affecting both human and natural capital. For example, the rapid expansion of bauxite mining has resulted in areas being turned into large red wastelands with several cases of bauxite runoffs caused by heavy downpours and “bauxite washing” processing generating effluents with traces of heavy metals, such as arsenic, mercury and aluminium, especially at Bukit Goh and Kuantan (Abdullah et al., 2016; Academy of Sciences Malaysia, 2017). According to a report, the level of air pollution from the bauxite dust recorded by monitoring the 24-h PM10 levels in Bukit Goh, Beserah (just outside residents' homes) and the Gebeng Industrial Estate, exceeded the standard levels of 150 μg/m³, under the 24-h Malaysian Ambient Air Quality Standard for PM10, by a record of 222.13 μg/m³ (Bukit Goh), 164.05 μg/m³ (Beserah) and 276.79 μg/m³ (Gebeng) (Naz Karim & Shah, 2016). This has led to the community suffering respiratory and skin rash problems due to daily exposure to bauxite dust (Abdullah et al., 2016; Academy of Sciences Malaysia, 2017). Hence, bauxite mining was halted for a period so that the government could improve the regulation of health impacts and ensure sustainable mining practices ahead, however, the strong demand for bauxite ore has meant the moratorium was lifted (Mazlan et al., 2019). These land-use conflicts and spill-over effects of the industrial and mining activities are expected to increase, and represent the types of complex integrated land-use challenges existing in the region, as noncomplementary land uses expand within closer proximity to each other, as shown by our model.

4.2 Simulating future land use and future research

The land-use change modelling portrayed how the historical land-use changes will lead to the future LULC regional pattern in 2030 (Figure 7). Figure 6 shows that oil palm plantations are projected to dominate about 41% (the equivalent of 283.61 km²) of the study area based on the current trend. The land-use change analysis also illustrated how different, often incompatible, land uses will be further concentrated in particular areas and grow in proximity to each other. Of concern for regional biodiversity, are potential impacts on existing forest cover, through a further reduction in overall area and fragmentation. The historical land-use change analysis suggests that conservation reserve status did not prevent deforestation, thus may not prevent future deforestation. Yet the forest reserves in this region are home to numerous protected species of mammals and birds (Chyuan, 2021). The conversion of forests will also likely result in secondary impacts through human-wildlife conflicts and increased risk of illegal encroachment leading to illegal hunting. While further industrialisation along the coast is likely to impact marine populations through unregulated pollution. The impact on natural ecosystems will cause a loss in ecosystem services, such as water regulation, especially for local residents (Lourdes et al., 2021).

The modelled 2030 projection provides insights into the compatibility and implications of high-impact land uses such as new industrial sites and mining in the vicinity of waterbodies, forest reserves, and adjacent to residential neighbourhoods. While the footprint of industry expansions was not precisely identified in the modelling, it was able to capture likely hotspots for industrial development (Figure 7). In regard to the quality of the simulated mine site locations from inland at Bukit Goh to coastal areas (Hartmann & Moosdorf, 2012), they do overlap with relevant underlying basic volcanic rock which is dominated by basalt lithology (Figure 4), the parent rock for bauxite deposits found in Kuantan (Iqhlima Najwa et al., 2019). One of the proposed mining areas simulated is under the Pahang State Development Corporation (PKNP) Mineral Industries and BG Mining operator surrounding the Bukit Goh area covering 3642 ha is expected to yield 170,000 tonnes of bauxite ore per month for consecutive 3 years (PKNP, 2019).

All forms of industrial, residential and mining land-use changes are likely to proceed 'business as usual' as approximated by the model in terms of spatial hotspots of change, though the total change in area may increase due to new development drivers not present in the historical data. In contrast to older, less profitable and more traditional industries, which provide fewer jobs (i.e., small-scale businesses, and agricultural and fisheries activities), mining and industrial sites are expected to expand greatly in the future. One of the key overall drivers is the ECRL project which will run along the ECER region connecting the west coast of Peninsular Malaysia and other main trade routes (Huam et al., 2018). The planned 620-km ECRL route will improve logistics between the Klang and Kuantan ports, reducing the shipping time, and enhancing connectivity between Malaysia and China (Lopez, 2016). The ECRL track within the study area includes three stations in Cherating, a major tourism spot, and two more stations near the Kuantan Port, for mass freight transportation such as mineral and agricultural commodities such as bauxite and palm oil products, and also manufacturing and chemical liquid goods, and marine produce (Malaysia Rail Link, 2019; Zainuddin, 2019) as well as tourism (ECER Master Plan 2.0, 2019; Loke, 2019). According to the Hong Kong Trade Development Council (2015) and the local district plan (Kuantan Municipal Council, 2015), the industrial park in Gebeng will expand to over 3500 acres making way for more high-end technology development, petrochemical and chemical manufacturing companies, and multipurpose development including light industry, commercial property and tourism parks. Large-scale infrastructure projects such as the expansion of the New Deep Water Terminal (NDWT) project and the expansion of Kuantan Port into a deepwater port will double the capacity of the port by 52 million freight weight tonnes, enabling larger ships to berth (Foon, 2017), which will also drive economic development in other sectors and therefore land-use change. Residential areas are also projected to increase in parallel, to keep pace with the population growth as the area develops with new job opportunities (ECER Master Plan 2.0, 2019).

In contrast to other land uses like industrial and mining development, the majority of oil palm plantation expansion may not be as extensive as predicted, due to the government's plan to cap the extent of Malaysian's palm oil estate at 6.5 million hectares from the year 2023 (Butler, 2019; Chain Reaction Research, 2021). The region's development vision may not be represented by historical changes as it shifts its focus towards the industry and commercial sectors, which are predicted to accelerate rapidly in the future (ECER Master Plan 2.0, 2019). However, our historical analysis showed that conversion to oil palm and deforestation took place even in areas that were set aside for conservation, and therefore complexities associated with on-the-ground responses to top-down planning, as well as the potential for changing policies need to be considered. Our future land-use change modelling represents a cautionary vision for Kuantan, and these land-use change footprints should be evaluated alongside other information sources, such as policy documents.

On the basis of the land-use change analysis validation, the accuracy of the land-use change maps and modelled projections; the outputs are fit for purpose, despite the potential for not accounting for future land-use policies and minor errors in the input data. Uncertainty in input data such as remote sensing maps is an ever present challenge for landscape ecology research (Lechner et al., 2012). To improve the spatial analysis and provide further insights, other GIS techniques such as public participation geographic information system (PPGIS) can be used to acquire local knowledge and perspectives at specific locations (Fagerholm et al., 2012). Ultimately, the land-use change analysis and modelling were able to depict the dynamics of land-use change in Kuantan following a “business as usual” trend, demonstrating that land-use change will likely result in significant impacts on the region's vulnerable environment.