3.1. Land Cover Change Analysis
The highest class recorded for 2009 was primary forests (
Figure 4) (578.15 km
2, 86.33% of total land area) followed by urban areas (50.10 km
2, 7.48%) and farm and mixed vegetables (33.00 km
2, 4.93%). The two minor land cover classes, namely cleared lands (7.84 km
2) and water bodies (0.60 km
2), were also recorded. The farm and mixed vegetables class includes tea plantations and other vegetation crops, such as corn, grapes, flowers, tomato, and strawberry farms. Within the farm and mixed vegetables land cover type, “other vegetation”, which comprises uses that are not individually categorized, account for only 1/10 of the total area. This is due to the lack of growth in area of these crop types in Cameron Highlands. Due to the limitations of Landsat 30 m resolution, these finer details are not able to be classified; hence, the difference can only be seen when higher resolution data are used.
In 2014, image classification results (
Figure 5 and
Table 4) show that the highest class was also primary forest (541.76 km
2, accounting for 80.90% of the total land area and representing a decline of 36.39 km
2 compared to 2009), followed by urban areas (59.50 km
2, 8.88%), and farm and mixed vegetables (57.03 km
2, 8.52%). The remaining land cover classes were cleared lands (11.06 km
2) and water bodies (0.34 km
2). Large changes across the terrain and hilly areas of Cameron Highlands can be seen within a five-year period. Due to the continuous rise in both urban areas (9.40 km
2) and farmland (24.03 km
2), the decline of primary forests continues. Massive deforestation has occurred in the study area based on image interpretation and classification.
Illegal land clearing in small communities also contributes to the overall problem of deforestation. Although this may represent a small change, it is nonetheless an issue for the Land and Planning authorities of Cameron Highlands. This is because the areas of these activities are deep within the rainforest. Moreover, such uncontrolled activities might also be the result of landowners breaching the agreement of forest clearing, which is undertaken for plantations rather than the agreed purpose of urbanization. This may occur in other similar mountainous sites in addition to Cameron Highlands [
49]. Predominantly in the areas of Kuala Terla, Tringkap, Brinchang, and Ringlet, in which numerous urban areas exist, it can be seen that farmlands are increasing in line with the decrease of urban housing. This could be due to the reduction in demand for housing areas and a rise in crop growth to meet the demand for western crops in the country. Towards the east, an area that was previously heavily forested has been cleared for land to build a road and a hydroelectric plant and is now categorized as an urban area. Land cover change during the study period is shown in
Table 4.
Rapid forest degradation continues to expand until 2019.
Figure 6 shows that the North East region of Cameron Highlands experienced land cover change due to construction of a road network, in addition to agriculture plots adjacent to the highway. Primary forest (518.71 km
2, accounting for 77.46% of total land area, a decline of 23.05 km
2 from 2014) was maintained as the highest class, followed by urban areas (61.10 km
2, 9.12%) and farm and mixed vegetables (84.61 km
2, 12.63%). The minor land cover classes of cleared lands and water bodies accounted for 1.10 and 4.17 km
2, respectively. A noticeable change, of −6.89 km
2, was found for the cleared land class, and a rise in water bodies of 0.76 km
2 (
Table 5). This change in water bodies is due to the operational status of the Ulu Jelai Hydroelectric dam. The trend of results obtained from this classification is consistent with the findings of Rendana et al. [
26], Mohammadi et al. [
50], and Razali et al. [
51].
The accuracy of the OBIA classification for 2019 was validated using a ground-truthing assessment of GPS points collected from the study area. The results are applied in
Figure 6. The producer’s accuracy value was 94.60% and the user’s accuracy value was 94.65%, resulting in an overall accuracy of 94.60%. Details of the confusion matrix table can be found in
Table 6. Notably, only the classification results of 2019 could be validated given the availability of GPS points, and images from other years were unable to be validated. Hence, for the years 2014 and 2009 (shown in
Table 7 and
Table 8, respectively), the confusion matrix was derived using Google Earth to validate the accuracy (see
Figure 7).
3.2. Land Surface Temperature Change Analysis Based on Landsat Satellite Data Compared to MODIS
High temperature values were concentrated in the populous towns of Brinchang, Ringlet, and Kuala Terla (
Figure 8, for 2009;
Table 9), with typical values ranging between 23 and 26 °C in the months of the first quarter. The results of LST retrieval from Landsat shows a similar range in temperature during April. The monsoon season peaks in the period from the end of January to the start of March, hence there was variability in temperatures during this period. High and low temperatures exhibited differences resulting from the cloud patterns and wind dynamics in the highland areas. As a result, it was difficult to obtain a consistent average for climate measures in the study area. Based on the results of Landsat alone (
Figure 8 and
Figure 9), it is difficult to ascertain the rise in temperatures relating to the deforestation pattern and the development of urban areas. Thus, the MODIS platform, in addition to air temperature data obtained from the Meteorological Department of Malaysia, was used as a guide to show the rising temperature trends.
The results are satisfactory with regards to a rising temperature, and this trend was observed after comparison with the multiple datasets. Although the rate of change is not as steep as previously thought, the change is clearly and steadily increasing. As shown in 2011, 2013, and 2019, due to cloud patterns and atmospheric conditions, it is difficult to comprehend the climate during this period; although other years show a better depiction of land surface temperature, this is a result of poor atmospheric correction effects [
17,
38,
52]. By comparison, although MODIS shows lower variability of data between months, there are no absolute highs or lows, resulting in a more accurate depiction of the climate in Cameron Highlands. In 2014, the Landsat 8 sensor measured a maximum temperature of 40.9 °C, whereas MODIS recorded a maximum of 31.1 °C. A summary of the data can be seen in
Table 9, which shows the relationship over time. Although the difference between the two sensors is significant, it is clear that the recorded temperature is correct due to the trend in rising temperatures.
3.3. Air Temperature against Landsat and MODIS.
Air temperature data were acquired from the Meteorological Department in Tanah Rata, Cameron Highlands. Maximum and minimum data for each month corresponding to the Landsat satellite data time frame are shown in
Table 9. According to the data values from Met Malaysia, the highest recorded temperature was 26.4 °C, and the lowest temperature recorded was 11.8 °C. A negative difference denotes a higher estimation of LST compared to the recorded temperature, whereas a positive difference indicates a lower estimation compared to the recorded value. In the urban environment, the air temperature may vary in a broader range due to the urban heat index (UHI). For instance, the months of March and April display a higher temperature as the radiative energy emitted by the surface of pavements, vegetation, urban areas, soil, and water is at a maximum. This is due to an increase in the human influx to the region of Cameron Highlands, primarily due to an increase of visitors, and the passing of the monsoon season, which manifests as dry spells after the period of heavy rainfall [
7,
9]. The data are not relevant to the whole region but only to certain areas in which these data were taken from the weather retrieving nodes, because it is difficult to ascertain the “correct” temperature value for the region overall [
53,
54,
55,
56].
The year 2014 shows the highest deviation across all LST datasets with a minimum of −2 °C and maximum of −15.4 °C. The second highest deviation in temperature occurred in 2019, in which the minimum and maximum deviations were 13.7 and −5.7 °C, respectively. By comparison, a deviation between Met Malaysia and MODIS was found for 2014, with a minimum of −10.2 °C and a maximum of −5.6 °C (see
Table 10). Clear skies are a major requirement for accurate LST retrieval. Thus, these unrealistic LST estimations are partly due to the erratic cloud coverage seen in 2011, 2013, 2018, and 2019 (
Figure 9). Because the base image was not trained with a specific emissivity value in specific areas, the resulting image is as shown in the figure [
12,
17,
18,
33,
35,
38].
3.4. Verification of LST
To ensure the correctness of the results, the LST must be verified. For this purpose, we obtained data derived from MODIS imagery and air temperature data from the Meteorological Department of Malaysia with links to the station in Tanah Rata, Cameron Highlands. Initially, MODIS was reclassified to the same resolution as the Landsat TIRS sensor, that is 100 m; it is a challenge to ascertain the coverage of the LST using a pixel size of spatial resolution of 100 meters. Other challenges of this method of verification relate to uncertainty in the data quality, namely the correct calibration of the image and apparent implications for the data due to the sensor of the satellite. Air temperature data obtained at ground level was used to address these challenges [
42].
Due to the vegetation cover of the study area, much of the vegetation canopy temperature was approximately equivalent to the surface air temperature [
24]. Hence, the Landsat and MODIS data for this particular area (primary forest) can be considered a good reference for the air temperature recorded by the weather station.
Table 11 shows various datasets which were subjected to a comparison of user predictions for the calculation of the required values. The LST values obtained from Landsat for 2009 to 2019 were averaged, and the associated Root Mean Square Error (RMSE) and bias values were derived. The reference LST was taken from MODIS for air temperature because this was considered to be accurate. It can be seen that the comparison of MODIS LST with local meteorological data has the lowest values relative to the other datasets. With an RMSE of 4.53 and 5.56 °C for maximum and minimum values, respectively, the results are reasonable and consistent, given the nature of the dataset at a spatial resolution of 100 m.
3.5. The LULC Effect on LST
Because the study area is covered mostly by vegetation, the area that is most stable in terms of LST changes across the map is primary forest. However, in many other areas, particularly towns, urban zones blend with secondary forests (parks, plantations, and areas of human-made vegetative recreation). These areas have a significant effect on the overall LST values because humans are carriers of heat and thus contribute to the anthropogenic heat distribution of a surrounding area. When this heat is massed, it tends to stick to nearby surfaces such as those of buildings. It is known that heat rises while cool air falls. Furthermore, the majority of buildings in the town areas of Cameron Highlands are living quarters with a minimum height of 10 stories and a maximum of 20 stories. The affected areas are Brinchang, Ringlet, and Tanah Rata, in which the population density is higher than that of other towns. These areas radiate an excess heat of 0.5 to 1.0 °C. Similarly, results from a study by Zhou et al. [
22] showed that a similar urban environment in Austin, Texas, experienced a rise in temperature of 0.45 to 0.90 °C. The extent to which terrain factors affect the variation of LST depends on the intensity of external solar radiation on a given day. In his study, Zhao notes that the parks and forested regions of Austin have a cooling effect, in which the heat sinks of the area absorb heat and then release cold air.
In contrast, areas without parks display signs of overheating as a result of heat release, surface albedo, and an irregular complexity of landscape structures. Results obtained in the current study are similar to those of a study conducted in Oslo by Venter et al. [
57] in which it was found that the temperature of the air was at least 0.6 °C higher in LST modelled using satellite imagery (in urban zones compared to the surrounding outer city). The authors used both Sentinel and Landsat as a means of comparison with meteorological data. The study clearly showed that the temperature of air increases through an impervious surface area paired with a decline of vegetation.
It has been found that the air temperature is at least 2–3 °C lower than the surface temperature depending on the altitude and surface type [
48]. This is more evident in urban areas in which there are no pure vegetation pixels. Hence, the air temperature will be slightly lower (by 2–3 °C as previously mentioned) than the land surface temperature because ground radiation conduction is influenced by air particles [
58,
59]. In other regions of the world, studies have shown the usage of MODIS to estimate air temperature. In Northwest China, Lu et al. [
60] conducted an experiment to determine the relationship between sensor temperature readings and on-site measurements. The results showed an RMSE between 2.39 and 3.05 °C; these results were dependent on seasonal and weather variability. Additionally, in the North Tibetan plateau of China, Zhu et al. [
61] compared MODIS with air temperature to derive an RMSE of 7.45 °C and a bias of 6.21 °C. The cause of these higher values was attributed to the presence of cloud pixels and haze surrounding the study area. In Egypt, El Kenawy et al. [
53] reported an average estimation of 5 °C according to seasonal change. These results are consistent with those of the present study, in which we derived RMSE values ranging between 4.53 and 8.13 °C for different sensor comparisons. Furthermore, Gomis-Cebolla et al. [
62] studied the use of LST from MODIS to compare with locally extracted air temperature values in the Amazon rainforest. The study reported RMSE values between 1.58 and 1.85 °C. Because the study was carried out during the winter, the expected RMSE was below estimation. By comparison, if the study was conducted during the summer, the RMSE would be expected to be between 4 and 5 °C.
Various studies of Asian cities comparable in size and density have observed an excess of anthropogenic heating as high as 1.5 °C [
21,
48,
63,
64,
65]. A study by Wang et al. [
66] conducted in Kuala Lumpur states that “air temperature is a good comparison against satellite-based data obtained temperatures, however, they do not reflect on the ‘true’ temperate climate of a surrounding area with a multitude of external implications such as wind speed, time, body induced heat radiation as well as heat created from vehicles”. This statement is also supported by the study of Kong et al. [
45], in which air temperature data also showed signs of outliers, mostly consisting of highs during the summer months of April to June, with estimated values of 0.3–0.5 °C. This could be a case of radiative heat being trapped amongst buildings or another case of climate-induced change as a result of increased influx into the region.
Cameron Highlands is known as a mountainous region with a climatology different from that of surrounding areas. Hence, multiple changes in temperature are frequent throughout the day. When paired with sudden changes in wind speed (which range on average from 2 to 3.4 kmph), this variation leads to uncertainty in the daily average temperature values. It should be noted that the undulating terrain of Cameron Highlands makes it difficult for heat to be radiated smoothly compared to the flatlands. This is more evident when various tall structures hinder the flow of air. Analysis of LST conducted in primary forests in urban areas reveal a higher pattern than LST of primary forests in a mixed zone of urban land cover. Portions of secondary forests are covered with glass roofing, whereas others have aluminum roofing. Furthermore, these roofs are either black in color or comprised of gray cloth to keep insects out. We classified these areas as secondary forests despite their signature showing as an urban zone due to surface reflectance. Notably, these areas have a similar pattern in their retention of heat to that of urban areas because plantations are situated below aluminum roofs [
67]. Thus, heat will further radiate in these areas, thus contributing to an increase in the value of the overall LST, because these manmade objects do not regulate LST in the same way as natural objects [
68,
69]. A more significant surface area density of buildings has a proportionate relationship to high urbanization and, in turn, a weaker LST regulation ability. This results in an increase in the massing of local heat, in addition to increases in the urban heat Index (UHI) and LST. Therefore, it is vital to take note of these particular areas that might affect the overall distribution of LST.