Differential analysis of landscape patterns of land cover products in tropical marine climate zones – A case study in Malaysia

2.1 Study areas

Malaysia is situated between the Pacific Ocean and the Indian Ocean and covers an area of 330,300 km², the whole of which is divided into Peninsular Malaysia (Peninsular) and the Sabah and Sarawak (Sasak) peninsulas by the South China Sea. The entire territory lies between 1–7°N and 97–120°E, making it one of the countries of South East Asia. Malaysia is one of the top producers and exporters of palm oil and rubber in the world and has access to abundant natural resources. Malaysia’s terrain is high in the north and low in the south, with the flat, undulating ground, and the Titiwangsa range, the peninsula’s primary mountain range, divides the peninsula into an east and west coast, with the interior being primarily hills and mountains and the coast of the sands being primarily plain. Malaysia is located near the equator and an average temperature between 26 and 30°C. Rainfall is abundant throughout the year, with annual precipitation ranging from 2,000 to 2,500 mm, up to 3,000 mm in West Malaysia, and up to 4,000 mm in East Malaysia. The research area’s geographic location is depicted in Figure 1.

Figure 1

Location of the study area.

2.2 Data

The GLOBCOVER, CCI LC, and MCD12Q1 land cover data were selected to analyze landscape pattern differences in the Malaysian region. The GLOBCOVER data are created by ESA in collaboration with the IGBP (International Geosphere-Biosphere Programme), UNEP (United Nations Environment Programme), and FAO (Food and Agriculture Organization) under the ESA-GlobCover project [35] (https://earth.esa.int/web/guest/data-access/catalogue-access). The European Space Agency’s CCI LC data are available annually from 1992 to 2015 (https://anon-ftp.ceda.ac.uk/neodc/esacci/land_cover/). The MCD12Q1 data are obtained by the NASA MODIS Center at Boston University and classified by using decision trees and neural networks (https://e4ftl01.cr.usgs.gov/MOTA/MCD12Q1.051/2013.01.01/). Table 1 provides the basic information of the three kinds of data.

Pre-processing of the three land cover data, including cropping, projection conversion, uniform resolution, and normalized classification system, is required prior to carrying out the comparative analysis. First, the study area land cover data were cropped by using ArcGIS 10.3 software with Malaysian vector boundaries. Second, the coordinates of all products were converted to the WGS 84 coordinate system and UTM projection. The data at different resolutions were then uniformly upscaled to 500 m using the nearest-neighbor resampling method in order to analyze the different characteristics of landscape patterns in three land cover data from the perspective of landscape ecology by selecting landscape indices that can quantitatively describe the distribution of landscape pattern. According to existing studies [36], data resampling may affect the accuracy of the original data, but it has little impact on the spatial pattern consistency study carried out in this study. Due to the various classification systems for the different classification products, the codes for the same types were standardized, and the overly detailed land cover types were grouped together for comparative analysis, resulting in a new, unified, and generalized classification system [37]. The three kinds of data after pre-processing are shown in Figure 2, and the correspondence between the new classification system and the original classification system (Table 2) of the three data is shown in Table 3.

Figure 2

Three kinds of land cover data pattern distribution.

3.1 Areal correlation analysis

The area of each type was aggregated for each of the three data types, and correlations were obtained for the corresponding land type area series between the different data, thus evaluating the area correlation between the different data [38]. The area correlation coefficient is calculated by the following formula:

where i denotes the i th land cover data combination; k denotes type; n denotes the number of types; X k and Y k denote the area of type k in data X and Y , respectively; and X ̅ and Y ̅ denote the average area of all land cover types in data X and Y , respectively.

3.2 Confusion matrix comparison

The confusion matrix is a popular tool for assessing a sample’s correctness. It is a matrix that shows how well the classification outcomes match the actual surface category [39]. The accuracy metrics available from the confusion matrix include (1) overall accuracy (OA), which tells the user the likelihood that a randomly chosen point in the classification result will be correctly classified and the percentage of data in the region that is correct. (2) User accuracy (UA), which is the proportion of mappings in a given category within a region that is correct, providing the user with a conditional probability of the actual classification being correct in a given region. (3) Producer accuracy (PA), which is the proportion of areas on the ground where a particular category is mapped to that category, representing the conditional probability that the ground truth category of that category is correctly classified. (4) Kappa coefficients, which are often used as an overall measure of accuracy. The formula for each indicator is as follows [40]:

where x ii is the correctly classified pixel number of type i ; n is the total pixel number; x i + is the total pixel number of type i in the data to be verified; x + i is the total pixel number of type i in the reference data; r the number of rows in the confusion matrix; and N is the total number of sample points.

Due to its precise placement, extensive coverage, high resolution, quick accessibility, and rich time phase, Google Earth imagery is one of the primary data sources for accuracy evaluation [17,41]. In order to reduce the negative impact of positioning and interpretation errors on sample quality, the following principles should be followed when selecting and interpreting samples: (1) To reduce the effect of a positioning error, the sample points were chosen as the center of a homogeneous area of the size of the error range area. (2) For some of the more difficult samples to interpret, the interpretation is assisted by combining references to other information. For example, combining other map information to distinguish between bare ground and artificial ground, using the voluntary geo-information platform Geo-wiki to assist in the interpretation, etc. (3) Multiple independent interpretations were used, and the sample was discarded when the interpretation results could not be agreed upon after negotiation. Based on the above principles, a final sample of 688 validation samples covering the study area was obtained by visual interpretation (Figure 3).

Figure 3

Spatial distribution of samples in the study area.

3.3 Spatial pattern analysis based on landscape indices

Understanding the landscape’s spatial organization is essential to understanding its dynamics and function. In landscape ecology, landscape indices contain basic information about landscape patterns, but they can also reflect the structural components and distribution patterns of the landscape itself in a simple and quantitative way and are used more frequently in the quantitative study of landscape patterns [17]. Therefore, in order to compare the consistency of spatial patterns, this article draws on previous research findings [42] and chooses three landscape indices that can reflect the characteristics of the landscape pattern in the study area. This comparison of the three data is made at both the macro and micro levels [43]. The selected landscape indices are as follows:

1. Largest patch index (LPI). This is the percentage of the overall landscape area that the largest patch of a certain patch type occupies. Changes in the LPI value can alter the intensity and frequency of disturbance, reflecting the direction and severity of human activity. The size of the LPI value impacts the biological characteristics of the landscape, such as the abundance of dominating and interior species. The calculation formula is as follows:

   (6) LPI = M A × 100 ,

   where M is the area of the largest patch of a given type of patch and A is the total landscape area. Values range from 0 < LPI ≤ 100 .

2. Number of patches (NP). It describes the heterogeneity of the whole landscape and the magnitude of its value is positively related to the fragmentation of the landscape. The formula is calculated as follows:

   (7) NP = n i ,

   where n i is the NP of landscape type i , taking values in the range NP ≥ 1 .

3. Landscape Shape Index (LSI). It reflects the complexity of the overall landscape shape, with values closer to 1 indicating a simpler overall landscape shape. As the patch types become more discrete, the LSI will gradually become larger and has no maximum limit. The formula is as follows:

where E is the total length of all patch boundaries in the landscape and A is the total area, taking values in the range LSI ≥ 1 .

4.1 Type composition correlation

Area is an important piece of information embedded in land cover data, and it is more relevant to compare the area of the three data. Figure 4 displays the area distribution in Malaysia for the three products. The results indicate that MCD12Q1, GLOBCOVER, and CCI LC land cover data all show that forest is the dominant land cover type, with 81.78, 54.57, and 57.05% of the area, respectively. The greatest variation in grassland/shrubland types for each land cover type was found in the MCD12Q1, GLOBCOVER, and CCI LC data, with 0.94, 1.21, and 0.01% of the area, respectively. Consistency between the MCD12Q1 and GLOBCOVER data is high for grassland/shrubland and wetland types and low for other types. The area agreement between the MCD12Q1 and CCI LC data was high for the artificial surface types and low for the other types. The consistency of area between GLOBCOVER and CCI LC data was high for forest, cropland, and water types and low for other types.

Figure 4

Land cover type construction diagram.

Table 4 shows the type area correlation coefficients for the three types of data. The area correlation coefficient between the GLOBCOVER and CCI LC data had the highest agreement, with a correlation coefficient of 0.998. The area correlation between GLOBCOVER and MCD12Q1 data was the least, with a correlation coefficient of 0.858. The findings of the consistency analysis of the area correlation coefficients for the aforementioned three kinds of land cover data agreed with those of the area share results in Figure 4. The reason for this is that both GLOBCOVER and CCI LC data use the LCCS classification system with 22 total categories, while the MCD12Q1 product uses the IGBP classification system with 17 total categories, and the difference in the original classification system may lead to a lower consistency between the MCD12Q1 product and the other two products.

4.2 Accuracy analysis based on confusion matrix

Figures 5 and 6 show the absolute accuracy of the three data obtained by validating the sample points. The CCI LC data had the highest OA and kappa coefficient of 59.01% and 0.4957, respectively, while the GLOBCOVER data had the lowest OA and kappa coefficient of 49.24% and 0.3614, respectively. From the analysis of the different types of PA, the consistency between GLOBCOVER and CCI LC data is high for cropland, forest, and water types, especially for the cropland, where the difference in PA between GLOBCOVER and CCI LC is about 1%. The consistency of wetland-type PA between the MCD12Q1 and CCI LC data was high, at 48.81 and 50.00%, respectively. From the analysis of different types of UA, the consistency of cropland UA between GLOBCOVER and CCI LC data is high, 34.64 and 39.32%, respectively, whereas other varieties have poor consistency. The consistency of UA between MCD12Q1 and CCI LC data is high for artificial surfaces and low for other types of UA.

Figure 5

OA and Kappa coefficient for three data.

Figure 6

PA and UA for three land cover data. 1 is Cropland; 2 is Forest; 3 is Grassland/Shrubland; 4 is Wetland; 5 is Water; 6 is Artificial surfaces; 7 is Bareland.

4.3 Overall spatial pattern comparative based on landscape indices

A longitudinal comparison of NP, LPI, and LSI was made for different land cover types (Figure 7). For the NP values, there is little difference between the GLOBCOVE and MCD12Q1 product for the grassland type, little distinction between the CCI LC and GLOBCOVE product for the water type, little distinction between the CCI LC and MCD12Q1 product for the bareland type, and a large distinction between the other types of NP values. For LPI values, there is little distinction between the CCI LC and GLOBCOVE products for the forest, wetland, and water categories and between the CCI LC and MCD12Q1 products for the artificial surfaces type. For the LSI values, there is little distinction between the forest, artificial surfaces, and bareland types for the CCI LC and MDC12Q1 products, little distinction between the grassland types for the GLOBCOVE and MCD12Q1 products, and little distinction between the water types for the CCI LC and GLOBCOVE products. As a result, the overall landscape pattern is characterized by a high consistency of the water landscape index between the CCI LC and GLOBCOVE products, a high consistency of the artificial surfaces landscape index between the CCI LC and MCD12Q1 products, and a high consistency of the grassland/shrubland landscape index between the GLOBCOVE and MCD12Q1 products.

Figure 7

Landscape pattern index comparison. 1 is Cropland; 2 is Forest; 3 is Grassland/Shrubland; 4 is Wetland; 5 is Water; 6 is Artificial surfaces; 7 is Bareland.

4.4 Local spatial pattern comparative based on landscape indices

In the NP landscape index spatial pattern distribution of a 10 km grid (Figure 8), the CCI LC and MCD12Q1 products have a higher consistency of landscape pattern in the East Malaysia region, and all three products have a lower consistency of landscape pattern in the West Malaysia region, which also indicates that the accuracy of land cover product classification is more difficult to ensure in areas of high landscape fragmentation and producers and users of different products should focus on this area.

Figure 8

NP landscape index comparison at the 10 km scale.

In the 10 km grid landscape index spatial (LPI) pattern distribution (Figure 9), the higher consistency of the landscape pattern in the southern half of the East Horse area for the three products suggests the presence of the largest patch of a particular patch type occupying a larger proportion of the overall landscape area in the area, resulting in a higher consistency of the landscape pattern in the area. According to the spatial distribution of different land cover products in Section 2.2, the area of forest types in these areas with high consistency accounts for a large proportion of the total area of the study area. The consistency of landscape patterns of all products was low in the West Malaysia region.

Figure 9

LPI landscape index comparison at the 10 km scale.

In the 10 km grid LSI landscape index spatial pattern distribution (Figure 10), the CCI LC and MCD12Q1 products show high consistency in the landscape pattern in the East Malaysia region and maintain high consistency in the central part of the West Malaysia region, while the GLOBCOVER product shows low consistency with both other products, indicating that the CCI LC and MCD12Q1 products’ landscape complexities are largely consistent. The findings show that the landscape pattern between the product spaces is less consistent with the more complicated landscape form of the land cover type.

Figure 10

LSI landscape index comparison at the 10 km scale.

5.1 Influencing factors of inconsistent accuracy of different land cover products

The change in land cover in tropical marine climate zones, which has the biggest biological gene pool on Earth, has a direct impact on the global environment. A remotely sensed data mapping system enables data support for research in this field. The following factors may contribute to the low consistency amongst the three global land cover products that were subjected to a more thorough evaluation and study in this article.

1. One of the main causes of inconsistent classification findings is the use of different classification systems [44]. The worldwide land cover classification system was created with full consideration of the characteristics of global land cover, it unavoidably leads to restrictions on its applicability to specific geographic places, such as tropical marine climate zones. This will also have an impact on the consistency of analysis across products [45]. The MCD12Q1 product includes only 17 categories, whereas the CCI LC and GLOBCOVER product classification system is more refined. However, there is variability in the definitions of certain types of these products in the classification system. For example, the GLOBCOVER product clearly defines broadleaved or deciduous shrubland from closed to open greater than 15% as grassland/shrubland in the various vegetation types of the tropical maritime climate zone, whereas the MCD12Q1 and CCI LC products only define the shrub category and do not explicitly give a value for vegetation cover. Because these vegetation categories are defined differently throughout the three products, there is little consistency in the varieties of grassland and shrubland. As a result, when constructing the classification system, a uniform value for the vegetation cover, tree height, and other factors should be clearly stated in order to eliminate the uncertainty that the classification system causes.

2. The quantity and quality of the evaluation sample can also add errors to the evaluation results [46]. To get a current sample set, collecting validation samples must be a team effort and involves a lot of work. Not only is the number of validation samples used in this article limited, but the article also does not take into account the possible uncertainty in the samples themselves, which will introduce some errors in the evaluation results [47]. In order to make it pertaining to the evaluation of different map data, i.e., to improve the applicability of the validation sample, future studies could investigate this issue based on Foody’s methodology [48]. To sustain the value of the validation sample, it is also crucial to acknowledge the significance of upgrading the sample set [49].

3. There is a great degree of consistency between goods for kinds with more unique spectral and textural properties (such as water and man-made surfaces). For tropical marine climate zones, various vegetation types (e.g., woodlands, shrubs, and grasslands) are susceptible to the formation of dissimilarities in the imaging process because of little variations in spectral and textural features and comparable living forms [50]. Optical remote sensing is more difficult to classify accurately, making the consistency between these confusing types low [51].

To sum up, producing land cover data products in tropical rainforest climate areas requires the formulation of appropriate classification systems according to the research purpose, significance, and regional characteristics. Besides, on the basis of improving the quality of classification sample points, multi-source data, including high-quality remote sensing images (such as low cloud cover) and some auxiliary data (such as geomorphic data), should also be integrated to improve the classification accuracy of land cover in this area.

5.2 The research method in this article is compared with existing methods

Table 5 shows several analysis results of land cover accuracy evaluation. The result of Gao et al. [52], through the area consistency and spatial consistency methods, concluded that the regions with inconsistent accuracy of three 30 m land cover products mainly occurred in heterogeneous regions. Hua et al. [53] conducted composition similarity analysis, confusion matrix analysis, and spatial multiple-consistency analysis, examined inter-annual changes, and concluded that the spatial consistency of Europe is high. In addition, the overall consistency in the EF climatic zone is very high. The surface conditions and data producers affect the spatial consistency of land-cover datasets to different degrees. CCI LC and GLCNMO (2013) have the highest overall consistencies on the global scale. Liang et al. [29] evaluate the accuracy of such products by using two sets of sample points collected from the Arctic region and concluded that GlobeLand30 and CCI-LC do not vary much from each other in terms of OA. They differ the most in the classification accuracy of shrub-covered land; CCI-LC performed better than GlobeLand30 in the classification of shrub-covered land, whereas the latter obtained higher accuracy than the former in the classification of urban areas and cropland.

Predecessors evaluated the accuracy of land cover products through different analysis methods, and all found the reasons for the consistency and inconsistency of land cover accuracy. However, they rarely carried out microscopic interpretation from the perspective of ecology on the scale of the grid. The consistency analysis of land cover products from a microperspective can better guide and optimize the strategies of automatic land cover classification such as zoning, stratification, and sample selection.