**2. Materials and methods**

#### **2.1. Study area**

measures have been adopted, but it is still difficult to prevent the loss of natural tropical for-

In China, the total area of tropical forest is currently increasing because of national reforestation policies, but the proportion of natural tropical forest to economic forest has changed in recent years [5]. As a result, changes have occurred affecting the maintenance of biodiversity, ecological functioning, and stability of forest ecosystem [6–8]. The Hainan Island of China has a large area of tropical forest, which serves as a natural treasure of biodiversity [2]. In recent years, the island has seen rapid economic development. Natural tropical forest is being constantly replaced by economic forest because of increasing demand for rubber, timber, and other forest products since 1988. To develop the economy in a sustainable and ecologically sensitive manner, the local government has proposed various ecological protection measures since 1998. In fact, the tropical forest in Hainan Island has greatly changed because of economic development and protection policies after 1998. Tropical forests are known to be impacted by population growth, economic development, national policies, and natural factors (such as terrain, climate, etc.) [5]. However, the key factors causing the changes of tropical

Because the proportion of natural tropical forest to economic forest has changed, two key challenges arise: (1) understanding how and why natural tropical forests are changing, and which factors have led to the changes; (2) understanding the fate and implications of the succession of different forest types. Thus, monitoring the dynamics of those changes occurring in natural tropical forest and economic forest becomes necessary, together with the identification of the main factors leading to those changes over broad spatial and temporal

Brandt et al. recognized that mapping forest distribution and succession are an essential component of forest biodiversity assessment [2]. Remote sensing provides an efficient technique for the monitoring and managing of tropical forests [9]. In addition, a combination of remote sensing and GIS techniques could help scientists discover the intrinsic forces driving the

However, some challenges remain in using remote sensing image classification. For example, for vegetation classification, there is a problem with mixed pixels resulting from same objects exhibiting different reflectance at varying wavelengths [10]. Also, it is difficult to improve the precision of the process of extraction without the support of a prior knowledge, such as the

In addition, the use of pixel-based methods is tiring and labor-intensive work, and misclassification of pixels is likely to occur because of errors during spectrum analysis [11]. For example, some natural forests lying within shaded areas of mountains tend to be regarded as other land use types. An object-oriented method could segment remote sensing images into different patch sizes based on integrated features of the spectrum, texture, shape of patch, and so forth [12]. In addition, the decision tree method could gradually extract individual land use types, using remote sensing extraction models and relevant auxiliary data, such as the distribution

ests due to conflicts between economic development and forest protection.

forest remain unclear.

116 Tropical Forests - New Edition

dynamics of forests.

of various forest types.

spatial distribution of various forests or patch shapes.

scales.

The study area covers 14,000 km2 in the central part of Hainan Island (**Figure 1**). The study area has a warm and humid tropical monsoon climate with annual average temperatures ranging from 22 to 26°C. Mountainous areas surround Five Finger Mountain and Yinggeling in Hainan Island. The island's natural tropical forest mainly includes tropical monsoon forest, tropical rain forest, evergreen broad-leaved forest, and coniferous forest. Economic forest mainly includes rubber plantations and orchards, which usually occur in the flatlands at the

**Figure 1.** The location of the study area (the central Hainan Island in China).

foot of the mountains. Some national and provincial reserves have been established since the 1970s to protect the biodiversity of these tropical forests.

The natural vegetation of Hainan Island is tropical monsoon forest, tropical rain forest, evergreen broad-leaved forest, mangroves, coniferous forests, shrubs, and grassland types. The main species of artificial vegetation are casuarina equisetifolia, eucalyptus, rubber, lemongrass, pepper, mango, and banana. The crops mainly include rice, sugarcane, sweet potato, cassava, and vegetables.

### **2.2. Data sources**

Landsat TM images (path/row 124/47) cover the study area. We used six images captured from December to the following February in 1988, 1998, and 2008; these images were obtained from the University of Maryland website (http://glcfapp.glcf.umd.edu:8080/esdi/ index.jsp) and satellite ground stations in China. The images were georeferenced with a precision better than 0.4 pixels. Finally, the digital data of these images were calibrated to surface reflectance values using the Fast Line-of-Sight Atmospheric correction of Spectral Hypercubes (FLAASH) Module in ENVI v.4.6. Other sources of information included a digital elevation model (DEM), and social, economic, and field survey data. DEM data were obtained from Aster satellite data (https://wist.echo.nasa.gov/~wist/api/imswelcome/). Slope and elevation data were derived from the DEM using the ARCGIS9.3 software. Some basic data (transportation corridors, population, national and provincial reserves, meteorological, social and economic data) were obtained from the Hainan Provincial Academy of Environmental Sciences. Field surveys were also conducted to collect information on the distribution of forests.

water index, which is a simple graphical indicator that can be used to analyze whether the target being observed is water or not. The vegetation index RVI is very sensitive to vegetation canopy chlorophyll content [15]. VIUPD is a vegetation index and sensitively reflects the amount of vegetation and the degree of vegetation vigor. GSI is grass and shrub differing index and can be used to identify bushes and grass from other kinds of trees [5, 15]. Then, the object-oriented forests information was extracted based on decision tree. The extraction results are shown in **Figures 4**–**7**. Multisource data (including slope from the DEM, pulp plantation planning maps, and other auxiliary data) were used to extract different land use types. According to the stage order, we first separated water-related ground objects (water bodies, paddy fields, and aquaculture areas) from images using the MNDWI model and

**Figure 2.** Examples of how phenology was used to discriminate the different forest classes. Representative pixels from three forest and other land use classes look similar under visual inspection on Landsat images from December to the

South China Tropical Forest Changes in Response to Economic Development and Protection…

http://dx.doi.org/10.5772/intechopen.73296

119

following February.

**Figure 3.** The steps to extract object-oriented information of forests based on decision tree.

#### **2.3. Methods**

#### *2.3.1. Extraction of multiple forest information datasets*

We compared the spectral differences of Landsat TM images captured in 1988, 1998, and 2008 of multiple forests in different seasons within a year, and selected images taken between December and the following January (e.g., December 1988 to January 1989) (**Figure 2**). Then we established flowcharts and processes for use with a sophisticated object-oriented decision tree method. First, these images were divided into multiscale segmentations based on the texture, spectrum, patch shape, and distribution of land use types using the Cognition 7.5 software. There are also other parameters need to be set, such as scale, shape, and compactness, which are obtained from the ground comparative research. Second, a set of indices extracting remote sensing information were calculated (**Figure 3**), including MNDWI = (Green – MIR)/ (Green + SWIR) [14], the Universal Pattern Decomposition Method (VIUPD), RVI = NIR/Red and the Grass and Shrub Differing Index GSI = (MIR–NIR)/(MIR + NIR). The wave bands of Green, Red, MIR, SWIR, and NIR were used in these models, since they contribute to identify different objects. It is concluded that the MNDVI, VIUPD, RVI, and GSI performed well in the information extraction of complex features. MNDWI is a modified normalized difference South China Tropical Forest Changes in Response to Economic Development and Protection… http://dx.doi.org/10.5772/intechopen.73296 119

foot of the mountains. Some national and provincial reserves have been established since the

The natural vegetation of Hainan Island is tropical monsoon forest, tropical rain forest, evergreen broad-leaved forest, mangroves, coniferous forests, shrubs, and grassland types. The main species of artificial vegetation are casuarina equisetifolia, eucalyptus, rubber, lemongrass, pepper, mango, and banana. The crops mainly include rice, sugarcane, sweet potato,

Landsat TM images (path/row 124/47) cover the study area. We used six images captured from December to the following February in 1988, 1998, and 2008; these images were obtained from the University of Maryland website (http://glcfapp.glcf.umd.edu:8080/esdi/ index.jsp) and satellite ground stations in China. The images were georeferenced with a precision better than 0.4 pixels. Finally, the digital data of these images were calibrated to surface reflectance values using the Fast Line-of-Sight Atmospheric correction of Spectral Hypercubes (FLAASH) Module in ENVI v.4.6. Other sources of information included a digital elevation model (DEM), and social, economic, and field survey data. DEM data were obtained from Aster satellite data (https://wist.echo.nasa.gov/~wist/api/imswelcome/). Slope and elevation data were derived from the DEM using the ARCGIS9.3 software. Some basic data (transportation corridors, population, national and provincial reserves, meteorological, social and economic data) were obtained from the Hainan Provincial Academy of Environmental Sciences. Field surveys were also conducted to collect information on the

We compared the spectral differences of Landsat TM images captured in 1988, 1998, and 2008 of multiple forests in different seasons within a year, and selected images taken between December and the following January (e.g., December 1988 to January 1989) (**Figure 2**). Then we established flowcharts and processes for use with a sophisticated object-oriented decision tree method. First, these images were divided into multiscale segmentations based on the texture, spectrum, patch shape, and distribution of land use types using the Cognition 7.5 software. There are also other parameters need to be set, such as scale, shape, and compactness, which are obtained from the ground comparative research. Second, a set of indices extracting remote sensing information were calculated (**Figure 3**), including MNDWI = (Green – MIR)/ (Green + SWIR) [14], the Universal Pattern Decomposition Method (VIUPD), RVI = NIR/Red and the Grass and Shrub Differing Index GSI = (MIR–NIR)/(MIR + NIR). The wave bands of Green, Red, MIR, SWIR, and NIR were used in these models, since they contribute to identify different objects. It is concluded that the MNDVI, VIUPD, RVI, and GSI performed well in the information extraction of complex features. MNDWI is a modified normalized difference

1970s to protect the biodiversity of these tropical forests.

cassava, and vegetables.

distribution of forests.

*2.3.1. Extraction of multiple forest information datasets*

**2.3. Methods**

**2.2. Data sources**

118 Tropical Forests - New Edition

**Figure 2.** Examples of how phenology was used to discriminate the different forest classes. Representative pixels from three forest and other land use classes look similar under visual inspection on Landsat images from December to the following February.

water index, which is a simple graphical indicator that can be used to analyze whether the target being observed is water or not. The vegetation index RVI is very sensitive to vegetation canopy chlorophyll content [15]. VIUPD is a vegetation index and sensitively reflects the amount of vegetation and the degree of vegetation vigor. GSI is grass and shrub differing index and can be used to identify bushes and grass from other kinds of trees [5, 15]. Then, the object-oriented forests information was extracted based on decision tree. The extraction results are shown in **Figures 4**–**7**. Multisource data (including slope from the DEM, pulp plantation planning maps, and other auxiliary data) were used to extract different land use types. According to the stage order, we first separated water-related ground objects (water bodies, paddy fields, and aquaculture areas) from images using the MNDWI model and

**Figure 3.** The steps to extract object-oriented information of forests based on decision tree.

**Figure 4.** Extraction of orchard information based on GSI and R5-R6 method.

**Figure 5.** Shrub information extraction based on GSI method.

then identified areas without vegetation coverage, such as urbanized areas, using the VIUPD method. Next, the RVI model, slope from DEM, and economic forest planning map were used to extract and delineate pulp plantation areas from other land use types. Finally, we separated natural forest and rubber plantation areas from grassland and orchards using the Grass and Shrub Differing Index (GSI) model and slope data from DEM. To correct some mistakes, we did some field investigation. For example, we initially were not sure whether some patches were rubber plantation or not in Landsat TM image, but we found through field investigation that these patches indeed were rubber plantation if they were near river and residential area, so we might correct these mistakes through water system distribution map.

*2.3.2. Change detection and key driving factor identification*

**Figure 7.** Rubber information extraction based on RVI and texture information method.

**Figure 6.** Extraction of pulp plantations information based on RVI method.

Three forest types and other land use types in the study area were extracted: natural forest, rubber plantation, pulp plantation, and other land use types (e.g., paddy fields, dry lands, orchards, sand, and urbanized areas). Map overlays of forest type for 1988, 1998, and 2008 were created using the GIS software to quantify the dynamic transformation between different forest types. Economic activities and protection measures were considered, and a comparative analysis method was used to compare natural rubber demand and price changes, the area of natural forests and the area of economic forests before and after the implementation of protection measures. To reveal the intrinsic driving factors from 1988 to 2008, in addition to considering economic and policy factors, the following factors were also considered: nature protection areas, farms,

South China Tropical Forest Changes in Response to Economic Development and Protection…

http://dx.doi.org/10.5772/intechopen.73296

121

South China Tropical Forest Changes in Response to Economic Development and Protection… http://dx.doi.org/10.5772/intechopen.73296 121

**Figure 6.** Extraction of pulp plantations information based on RVI method.

**Figure 7.** Rubber information extraction based on RVI and texture information method.

#### *2.3.2. Change detection and key driving factor identification*

then identified areas without vegetation coverage, such as urbanized areas, using the VIUPD method. Next, the RVI model, slope from DEM, and economic forest planning map were used to extract and delineate pulp plantation areas from other land use types. Finally, we separated natural forest and rubber plantation areas from grassland and orchards using the Grass and Shrub Differing Index (GSI) model and slope data from DEM. To correct some mistakes, we did some field investigation. For example, we initially were not sure whether some patches were rubber plantation or not in Landsat TM image, but we found through field investigation that these patches indeed were rubber plantation if they were near river and residential area, so we might correct these mistakes through water system distribution

**Figure 4.** Extraction of orchard information based on GSI and R5-R6 method.

120 Tropical Forests - New Edition

**Figure 5.** Shrub information extraction based on GSI method.

map.

Three forest types and other land use types in the study area were extracted: natural forest, rubber plantation, pulp plantation, and other land use types (e.g., paddy fields, dry lands, orchards, sand, and urbanized areas). Map overlays of forest type for 1988, 1998, and 2008 were created using the GIS software to quantify the dynamic transformation between different forest types. Economic activities and protection measures were considered, and a comparative analysis method was used to compare natural rubber demand and price changes, the area of natural forests and the area of economic forests before and after the implementation of protection measures.

To reveal the intrinsic driving factors from 1988 to 2008, in addition to considering economic and policy factors, the following factors were also considered: nature protection areas, farms, transportation corridors, and the elevation and slope of each site. The data of nature protection areas were derived from the map of national nature protection area; the data of farms and transportation were extracted from Landsat TM images; the data of elevation and slope were extracted from DEM. The study area was divided into grid cells, and the information for each grid cell was extracted by the spatial analysis model of the GIS software. By using random permutation tests [16] to choose grid cells in each year, the association of different forest types with the abovementioned factors was examined. We used the average Euclidean distance between the multiple forest patches and these driving factors to capture differences in conservation value and human activity. For multiple forest conversions, the following method was used to identify the key factors: the patch of the different forest area serves as the basic unit; then, adjacent patches with the same change trends were classified into a uniform block. The average value of a factor in each block with the same trend was used to serve as a sample, and the value of the factor served as a collective value. The ratio of the area covered by a particular land use type in 2008 was compared with the corresponding blocks of the two previous periods (1988 and 1998) to detect changes using the overlay. Then, we took the altitude, slope, and the minimum distance from the farmland and road to the area where transformation occurred as variable and took the transformation of forest types as dependent variable. General regression analyses between them were conducted by collectively analyzing the blocks with the same trends.
