Remote Sensing of Tropical Rainforest Biomass Changes in Hainan Island, China from 2003 to 2018

: The largest area of tropical rainforests in China is on Hainan Island, and it is an important part of the world’s tropical rainforests. The structure of the tropical rainforests in Hainan is complex, the biomass density is high, and conducting ground surveys is difﬁcult, costly, and time-consuming. Remote sensing is a good monitoring method for biomass estimation. However, the saturation phenomenon of such data from different satellite sensors results in low forest biomass estimation accuracy in tropical rainforests with high biomass density. Based on environmental information, the biomass of permanent sample plots, and forest age, this study established a tropical rainforest database for Hainan. Forest age and 14 types of environmental information, combined with an enhanced vegetation index (EVI), were introduced to establish a tropical rainforest biomass estimation model for remote sensing that can overcome the saturation phenomenon present when using remote sensing data. The ﬁtting determination coefﬁcient R 2 of the model was 0.694. The remote sensing estimate of relative bias was 2.29%, and the relative root mean square error was 35.41%. The tropical rainforest biomass in Hainan Island is mainly distributed in the central mountainous and southern areas. The tropical rainforests in the northern and coastal areas have been severely damaged by tourism and real estate development. Particularly in low-altitude areas, large areas of tropical rainforest have been replaced by economic forests. Furthermore, the tropical rainforest areas in some cities and counties have decreased, affecting the increase in tropical rainforest biomass. On Hainan Island, there were few tropical rainforests in areas with high rainfall. Therefore, afforestation in these areas could maximize the ecological beneﬁts of tropical rainforests. To further strengthen the protection, there is an urgent need to establish a feasible, reliable, and effective tropical rainforest loss assessment system using quantitative scientiﬁc methodologies.


Introduction
The rapidly increasing global population, land-use change, and large-area forest fires have resulted in a downward trend in global biomass in recent years [1][2][3][4][5]. With global warming, forest biomass could effectively delay the increase in carbon dioxide, which has attracted much attention [6,7]. Tropical rainforests in low latitudes account for 59% of the global forest biomass carbon sequestration [8,9]. The study of carbon sequestration    Using ArcGIS 10.8 software, the database was established as follows: (1) Based on the daily observation data of 553 meteorological stations in Hainan Island from 2003 to 2018, a climate database (1 km resolution) consisting of annual average maximum temperature, annual average minimum temperature, and annual average rainfall was constructed by spatial interpolation. By providing a prototype of the dynamic climate data, a website is able to maintain local climate data in the local system, while another centralized website continuously updates the data through a public database and provides access to all website data [24]. (2) The Chinese soil science database was utilized (the reference before 1 January 2021) [25]. In this study, spatial interpolation was used to generate the soil types: pH − H 2 O, pH − KCl, total nitrogen, potassium, and phosphorus contents, and other soil databases (2 km resolution). (3) Based on SRTM data and Hainan administrative boundary, a DEM terrain database (90 m resolution) [26] was generated. (4) Combining the population statistics data of Hainan Island for 2003,2008,2013, and 2018 with land-use types, night light brightness, and residential density, the population data were distributed to 3.06 million grids according to the multiple-factor weight distribution method, and a human-disturbance database (1 km resolution) of population dynamic evolution was generated by spatial interpolation [27]. In this way, the population of a specific area could be estimated by adding the population involved in each grid area. ( [28], 133 permanent sample plots were screened out, and the 133 permanent sample plots contained 16,280 sample trees. The forest biomass data of 70 main tropical rainforest tree species (280 trees) were collected using the standard tree full digging method [29]. A tropical rainforest biomass database of permanent sample plots was generated using the where W denotes tropical rainforest biomass, D is DBH, H is tree height, and a, b are parameters [29]. According to this model, we measured the forest biomass of 133 permanent plots in 2003, 2008, 2013 and 2018. Therefore, a forest biomass sample plot database for Hainan Island was established. (7) Based on the forest age from China's National Forest Continuous Inventory data in Hainan (in 2018), the scope of the natural forest resource protection project in Hainan, manual visual interpretation data of Landsat historical remote sensing images, and a forest age database (1 km resolution) [30] were generated by spatial interpolation. There were 14 pieces of environmental information involved in this study, all from the tropical rainforest database mentioned above ( Figure 4). Among them, the annual average maximum temperature ( • C), annual average minimum temperature ( • C), and annual average rainfall (mm) were from the climate database. The information on soil pH − H 2 O, soil pH − KCl, total nitrogen content of the soil (%), total potassium content of the soil (%), and total phosphorus content of the soil (%) were all obtained from the soil database. The slope gradient ( • ), slope direction ( • ), and altitude (m) were obtained from the DEM terrain database. The human-disturbance factor (person/ha) was obtained from the human-disturbance database. Latitude ( • ) and longitude ( • ) were used for geographical registration between different data. In addition, tropical rainforest biomass was derived from the tropical rainforest biomass database. Forest age was taken from forest age database.

Enhanced Vegetation Index Extraction Method Based on MODIS Product Data
In this study, the EVI index was calculated using MODIS product data provided by NASA [31]. The time and spatial resolutions were 16 days and 250 m [32], respectively, and the time series were 2003, 2008, 2013, and 2018. The product data of the study area covered by the image in each period includes two tile data points (h28v06 and h28v07). In addition, the vector image of Hainan was used to calculate a mask of the original image data. Using the batch processing function of the MODIS Reprojection Tool (MRT) software, the concatenation and projection of MODIS datasets (UTM zone 49n, WGS-84) was completed [33]. IDL programming technology [34] was used for vector clipping and other batch processing of the study area. Of the different image denoising methods in TIMESET, such as S-G filtering, asymmetric Gaussian fitting, and double logistic fitting, we used the S-G filter to preserve the image details. Furthermore, the algorithm was not limited by the data of the image itself, which improved the applicability of smoothing filtering.
At present, the strategy of using remote sensing estimation methods to calculate vegetation coverage mainly involves regression models, unit data value decomposition models, and the most used vegetation index methods. As there is a significant correlation between vegetation coverage and vegetation index, the unit data value dichotomy model (including 1%, 99%, and 5%, 95%) and maximum value composite (MVC) could be used to estimate the vegetation coverage. In this study, a MVC method was adopted to superimpose image layers with different cycles for each year, and the unit data value of each raster was taken as the maximum value among multiple cycle image layers, thus forming a method of synthesizing image data [33]. The calculation formula is as follows: where EV I i is the maximum composite value of EVI in the i-th year, which was also the highest value of vegetation coverage. EV I i−j is the value of vegetation coverage of each unit data value in each cycle, and each cycle is 16 days. As a result, there are 23 images per year (1 ≤ j ≤ 23).

Establishment of the Tropical Rainforest Biomass Estimation Model for Remote Sensing in Hainan
Previous experiments, based on the commonly used tropical rainforest biomass remote sensing estimation linear model (Model 2), considered only EVI to estimate the tropical rainforest biomass: where a is the model fitting parameter, and b is the constant. W is the tropical rainforest biomass. Considering that EVI may be saturated in tropical rainforest biomass estimation with remote sensing, a tropical rainforest biomass estimation model for remote sensing (Model 3) with forest age was established: where a, b, and c are the model fitting parameters of forest biomass change with forest age, and t is forest age. d is the model fitting parameter of the impact of EVI on forest biomass. Finally, the environmental factors indirectly affecting vegetation distribution, forest structure, forest growth, and forest vegetation spectral characteristics were introduced to establish a tropical rainforest biomass estimation model for remote sensing: where c i is the model fitting parameter of the impact of environmental information on forest biomass, and x i is the environmental information, including latitude, longitude, average annual maximum temperature, average annual minimum temperature, annual average rainfall, slope gradient, slope direction, altitude, human-disturbance factor, soil pH − H 2 O, soil pH − KCl, and the total nitrogen, phosphorus, and potassium contents of the soil. The three tropical rainforest biomass estimation models established in this study were all realized using IBM SPSS Statistics 25 software.

Data Preprocessing before Tropical Rainforest Biomass Estimation Model Fitting
In this study, Model 2, Model 3, and Model 4 were tested for multicollinearity. Among them, Model 2 was a linear model. The tolerance value and VIF value were 1.000 and 1.000, respectively. Model 3 and Model 4 were nonlinear models. The VIF value and tolerance value of EVI of Model 3 were 1.024 and 0.977, respectively. The VIF value and tolerance value of forest age were also 1.024 and 0.977, respectively. In Model 4, a multicollinearity analysis was performed using EVI, forest age, and 14 environmental factors. The VIF values of these 16 factors were between 1.215 and 7.248, and the tolerance values were between 0.138 and 0.823. The VIF values of the three models were all less than 10, and the tolerance values were between 0 and 1. The results showed that there was no multicollinearity in the three models. When different feature vectors come together, as a result of their own expression, the absolute value of the small data is negligible compared to that of the big data. This study needed to normalize the extracted features vector, ensuring that each feature vector was treated equally by the classifier (  The total phosphorus content of the soil on Hainan Island was between 0% and 0.76%; 14. Total potassium content of the soil (%): The total potassium content of the soil in Hainan Island was between 0% and 4.40%. Table 1. Normalization formulas of the 14 environmental information variables.

Environmental Information Normalization Formula
Human-disturbance factor (person/ha) Total nitrogen content of the soil (%) Total phosphorus content of the soil (%) X P = P−P min P max −P min Total potassium content of the soil (%)

Fitting Results and Accuracy Evaluation of the Remote Sensing Estimation Model for Tropical Rainforest Biomass in Hainan
We used 399 fitting sample data points and 133 verification sample data points from 133 forest sample plots from phase 4 (in 2003, 2008, 2013, and 2018). Model 3 and Model 4 were developed and built based on 399 fitting sample data units. IBM SPSS Statistics 25 software was used to fit the tropical rainforest biomass estimation model for remote sensing. Combined with the forest age, environmental information, and remote sensing factor EVI, the goodness of fit significantly improved from Model 2 (R 2 < 0.100) to Model 3 (R 2 = 0.568), and finally to Model 4 (R 2 = 0.694). The estimated values, standard errors, and 95% confidence intervals of the parameters fitted by Model 3 are listed in Table 2.
The estimated values, standard errors, and 95% confidence intervals of the parameters fitted by Model 4 are listed in Table 3. The results show that the human-disturbance factor had the strongest correlation with the tropical rainforest biomass estimation model in remote sensing. It was much higher than other environmental information. In addition to human interference with environmental information, the annual average rainfall, total phosphorus content of soil, longitude and tropical rainforest biomass estimation model for remote sensing were also strongly correlated. Annual average minimum temperature, soil pH − KCl, and the tropical rainforest biomass estimation remote sensing model had a weak correlation. We used 133 verification sample data points to analyze the accuracy of Model 3 and

Spatial Pattern of Tropical Rainforest Biomass in Hainan
Based on the longitude and latitude of 5662 remote sensing estimation plots (Figure 3), EVI, forest age, and 14 main pieces of environmental information were extracted from the tropical rainforest database. The values of the 14 major environmental information variables collected from the databases of climate, topography, soil, and human-disturbance factors, and tropical rainforest biomass in 2003, 2008, 2013, and 2018, were estimated using the fitted tropical rainforest biomass estimation remote sensing model. A spatial pattern distribution of tropical rainforest biomass in Hainan was then generated using ordinary kriging spatial interpolation (Figure 6), with a spatial resolution of 2.0 km.
The kriging method was used to interpolate the tropical rainforest biomass and forestland area (Mg/ha) of each city and county in Hainan Island, and the total tropical rainforest biomass and forestland area of 18 cities and counties were calculated (shown in Figures 7 and 8). The administrative division of Hainan Island from 2003 to 2018 is shown in Figure 2, and the biomass distribution of the tropical rainforest on Hainan Island is shown in Figure 7. The tropical rainforest biomass distribution in Hainan's cities and counties was inconsistent. In this study, the biomass distribution of the tropical rainforest in Hainan Island was ranked. The tropical rainforest biomass above 10 Tg was divided into the first gradient, which included Ledong Li Autonomous County, Qiongzhong Li and Miao Autonomous County, Wuzhishan City, Sanya City, Baoting Li and Miao Autonomous County, Baisha Li Autonomous County, and Dongfang City. Cities and counties with tropical rain forest biomass above 5 Tg were divided into the second gradient, namely, Changjiang Li Autonomous County, Qionghai City, Wanning City, and Lingshui Li Autonomous County. Cities and counties with biomass below 5 Tg were divided into the third gradient, and they are Tunchang County, Wenchang City, Danzhou City, Chengmai County, Ding'an County, Haikou City, and Lingao County. The tropical rainforest biomass in all cities and counties increased annually, and Ledong Li Autonomous County had the fastest growth rate.
The administrative division of Hainan Island is shown in Figure 2, from 2003 to 2018, and the forested area of Hainan Island is shown in Figure 8. The forestland area is concentrated in the central part of Hainan. We classified the forestland area in different cities and counties of Hainan Island. The forestland area of the first gradient was more than 10 5 ha, which included Qiongzhong Li and Miao Autonomous County, Ledong Li Autonomous County, and Baisha Li Autonomous County. The cities and counties with a forestland area of more than 5 × 10 4 ha were divided into the second gradient, namely Sanya City, Wanning City, Qionghai City, Wuzhishan City, Dongfang City, Tunchang County, Baoting Li and Miao Autonomous County, and Changjiang Li Autonomous County.
The cities and counties with a forestland area of less than 5 × 10 4 ha were divided into the third gradient, namely Danzhou City, Chengmai County, Wenchang City, Lingshui Li Autonomous County, Ding'an County, Linggao County, and Haikou City. Except for Lingshui Li Autonomous County, Baoting Li and Miao Autonomous County, Qionghai City, Wanning City, and Ledong Li Autonomous County, the woodland areas of other cities and counties showed a downward trend. Among them, from 2003 to 2018, Qiongzhong Li and Miao Autonomous County was the most seriously damaged forestland area.    Figure A1 of Appendix A.

Feasibility Analysis of Introducing Environmental Information to Estimate Forest Biomass
The common forest biomass estimation methods with remote sensing include adjusting features, such as remote sensing color band, parameters, and texture ( Table 4). The choice of research methods has a significant effect on forest biomass estimation in the temperate zone. However, in tropical rainforests with complex forest structures, the effect of remote sensing estimation was not significant. Foody used a neural network to estimate tropical rainforest biomass in Borneo, Malaysia. However, the remote sensing data of some plots were affected by clouds or cloud shadows, resulting in a small number of samples [37]. Rödig used a forest model to estimate the change in tropical rainforest biomass in the Amazon, combined with remote sensing and field work to improve the accuracy. However, this method underestimated approximately 15% of biomass [38]. Sambatti used airborne InSAR to estimate Amazon's tropical rainforest biomass and found that the best results were obtained by using X-band and P-band InSAR [39]. Hansen used airborne laser scanning (ALS) to model tropical rainforest forest biomass in Tanzania, and it was effective under extreme conditions of tropical rainforests [40]. However, the high cost of lidar and airborne InSAR, as well as the lack of data in some periods and tropical regions, were not conducive to tropical rainforest biomass estimation in Hainan from 2003 to 2018.
The most commonly used vegetation information measures for forest biomass estimation with remote sensing have been NDVI and EVI. Motlagh used Spot-6 satellite images and a regression model to analyze the accuracy measurement results of NDVI, RVI, and TVI on the forest biomass of Helka forest in northern Iran and concluded that NDVI had the highest accuracy [41]. However, NDVI is easily affected by the forest canopy, soil, atmospheric molecules, and excessive forest biomass [42,43], especially in tropical rainforests, which is close to saturation due to the complex canopy structure [44][45][46]. Bhardwaj also confirmed that NDVI might not be a reliable method for estimating forest biomass carbon storage in Himalayan subtropical forests [1]. For improving the sensitivity of tropical rainforest biomass estimation, the EVI was selected in this study. EVI could minimize soil brightness by introducing background adjustment parameters and atmospheric correction parameters to increase sensitivity to differences in forest canopy density. Moreover, EVI introduced a blue-light band, which could reduce the influence of the atmosphere using the difference in aerosol scattering between the blue-and red-light bands [47]. EVI was better than NDVI when used to extract vegetation information from remote sensing images, and it could filter out impurities outside the vegetation information.
However, neither NDVI nor EVI could overcome the saturation phenomenon of forest biomass estimation using remote sensing. Shen studied the relationship between forest coverage and forest biomass and found no significant correlation between aboveground biomass and EVI and NDVI when the canopy was covered with high flower density [47], and the RMSE of forest biomass estimation was low. Anaya also proved that there was no significant correlation between EVI and tropical rainforest biomass, even when the parameters were adjusted [48]. Eckert used Pearson's correlation analysis and stepwise multiple linear regression analysis to determine the correlation between forest biomass and EVI. Based on the spectral and texture information of Worldview-2, the model was established, and the accuracy of the degraded forests was found to be higher than that of non-degraded forests [49]. This further proved that when the forest biomass density is high, the EVI would produce the saturation phenomenon in remote sensing estimation.
Propastin used NDVI to simulate the spatial relationship between vegetation and rainfall in central Sulawesi Island and concluded that the spatial change in the study area was caused by the difference in potential environmental factors (such as the vegetation composition, soil type, hydrology, and land use) caused by terrain diversity [50]. Therefore, this study considered introducing environmental information to estimate tropical rainforest biomass. In addition, Propastin added altitude when estimating forest biomass with the remote sensing multispectral vegetation index (VI), which improved the accuracy of the traditional GWR model [51]. This indirectly showed the feasibility of introducing environmental information into this study to improve tropical rainforest biomass estimation accuracy. In this study, only 14 main environmental factors, including climate, topography, soil, and human-disturbance factors, were considered for tropical rainforest biomass estimation. With an in-depth investigation of tropical rainforest resources and environment, the Hainan tropical rainforest database could be further improved by considering natural disasters and other environmental information. Therefore, future research should use a multimodal data fusion method to carry out feature fusion of all environmental information, such as meteorological fusion factors, topographic fusion factors, soil fusion factors, natural disaster fusion factors, and human-disturbance factors, before tropical rainforest biomass remote sensing estimation model fitting. This can avoid not only the over-fitting problem but also ensure further improvement in the accuracy of tropical rainforest biomass estimation with remote sensing.

Evaluation and Analysis of the Spatial Pattern of Tropical Rainforest Biomass in Hainan
In this study, the ordinary kriging interpolation method was used to perform the difference analysis, as shown in Figure 9. From 2003 to 2018, the tropical rainforest biomass in Hainan was mainly distributed in the central mountainous area and southern area of Hainan, including Qiongzhong, Wuzhishan, Baisha, Changjiang, Dongfang, Ledong, and Tunchang. Only a few tropical rainforests are distributed in the northern and coastal areas, and the proportion of non-forest land was large, especially in the coastal areas of Wenchang, Haikou, Ding'an County, Chengmai County, Lingao County, Sanya City, Ledong Li, Miao Autonomous County, and Wanning City, which are close to the coastline. It was found that human-disturbance has a strong correlation with biomass, which is mainly reflected in the destruction of natural forestland areas. Indeed, from 2003 to 2015, land utilization changed in Hainan Island; with the development of tourism, the conversion of cultivated land to construction land took place in coastal plain areas, and the conversion of woodland to farmland mainly took place in the east, northwest, and southeast of Hainan. For income, the local government converted farmland into forests in Hainan, with coconut forests in the northeast, eucalyptus forests in the west, and rubber forests in the northwest [52]. The tropical rainforests in northern Hainan were largely replaced by economic forests. However, owing to the strong carbon sequestration, the tropical rainforest biomass in Hainan showed an increasing trend. The forestland area fluctuated greatly every year, and the forestland area decreased in some cities and counties, which slowed further increases in the tropical rainforest biomass. As the tropical rainforests were challenging to repair, the ecological loss caused by the reduction of tropical rainforest areas needs to be further evaluated.
The year 2013 witnessed severe changes in the tropical rainforest areas in Hainan. The main reason was that Hainan concentrated on developing tourism and the real estate industry as its main economy during this period, which seriously damaged the coastal tropical rainforests. In addition, low-altitude natural forests were at risk of being replaced by economic forests, such as fruit trees [53]. However, in Wanning City, Lingshui Li and Miao Autonomous County, Wenchang City, Haikou City, and Lingao County the forestland area gradually recovered. Part of the reason for this was the establishment of nature reserves. From 2000 to 2010, the forest area in nature reserves showed an increasing trend, while the adjacent unprotected areas and wider unprotected areas experienced deforestation [54]. The development of real estate in these areas was another concern. For maintaining the overall ecological balance of Hainan, the local government increased the area of forested land, but the area of restoration was far less than that of destruction. In addition, the ecological loss caused by replacing the destroyed natural rainforest with artificial rainforests needs to be further evaluated.
To meet the ecological civilization model advocated by Hainan, cities and counties with low woodland areas and high damage need to consider afforestation and ecological restoration in non-woodland and parts of unused construction land. From the fitting parameters of the tropical rainforest biomass estimation remote sensing model, the areas with large annual average rainfall had fewer tropical rainforests. However, sufficient rainfall would make the carbon sequestration of tropical rainforest biomass more evident [55]. Local governments could prioritize afforestation and ecological restoration in areas with abundant rainfall. In addition, tropical rainforest biomass is generally proportional to biodiversity [38]. The areas with high tropical rainforest biomass were suitable for adding nature reserves to strengthen the protection of tropical rainforests further, for example, in the north of Sanya City, in the west of Baoting Li and Miao Autonomous County, in the south of Wuzhishan City, in the south of Tunchang County, and in the south of Ding'an County.

Conclusions
To overcome the saturation phenomenon of tropical rainforest biomass estimation with remote sensing and improve the accuracy of spatial distribution estimation of tropical rainforests in Hainan Island, this study established the Hainan tropical rainforest database and introduced a tropical rainforest biomass estimation model that uses EVI, forest age, and 14 environmental variables for remote sensing. Based on this model, the temporal and spatial patterns of the tropical rainforest biomass in Hainan Island from 2003 to 2018 were evaluated and analyzed. This was a breakthrough attempt to estimate the spatial distribution of tropical rainforests on a large scale. Considering that Hainan will soon become a global free-trade port, any future development trends should effectively balance economic development and environmental protection. Future work will consider further improving the Hainan tropical rainforest database and increasing environmental information to improve the accuracy of tropical rainforest biomass estimation. This study can lay the foundation for using scientific and technological tools in the decision-making process and management of tropical rainforests. In addition, despite the construction of artificial tropical rainforests, the area of tropical rainforests in Hainan was found to have decreased annually, making it difficult to reach the previous ecological level. Therefore, establishing an ecological loss evaluation system for tropical rainforests will be an important future research direction.  Data Availability Statement: Data available in a publicly accessible repository that does not issue DOIs Publicly available datasets were analyzed in this study. This data can be found here: http: //www.zixuanqiu.com/nd.jsp?id=24#_np=105_462 (accessed on 27 April 2021).

Conflicts of Interest:
The authors declare no conflict of interest.

Appendix A
Histograms showing the forest biomass of each city and county in Hainan Island, and line charts showing the forest area of each city and county in Hainan Island.