Integration of GF2 Optical, GF3 SAR, and UAV Data for Estimating Aboveground Biomass of China’s Largest Artificially Planted Mangroves

: Accurate methods to estimate the aboveground biomass (AGB) of mangroves are required to monitor the subtle changes over time and assess their carbon sequestration. The AGB of forests is a function of canopy-related information (canopy density, vegetation status), structures, and tree heights. However, few studies have attended to integrating these factors to build models of the AGB of mangrove plantations. The objective of this study was to develop an accurate and robust biomass estimation of mangrove plantations using Chinese satellite optical, SAR, and Unmanned Aerial Vehicle (UAV) data based digital surface models (DSM). This paper chose Qi’ao Island, which forms the largest contiguous area of mangrove plantation in China, as the study area. Several field visits collected 127 AGB samples. The models for AGB estimation were developed using the random forest algorithm and integrating images from multiple sources: optical images from Gaofen-2 (GF-2), synthetic aperture radar (SAR) images from Gaofen-3 (GF-3), and UAV-based digital surface model (DSM) data. The performance of the models was assessed using the root-mean-square error (RMSE) and relative RMSE (RMSEr), based on five-fold cross-validation and stratified random sampling approach. The results showed that images from the GF-2 optical (RMSE = 33.49 t/ha, RMSEr = 21.55%) or GF-3 SAR (RMSE = 35.32 t/ha, RMSEr = 22.72%) can be used appropriately to monitor the AGB of the mangrove plantation. The AGB models derived from a combination of the GF-2 and GF-3 datasets yielded a higher accuracy (RMSE = 29.89 t/ha, RMSEr = 19.23%) than models that used only one of them. The model that used both datasets showed a reduction of 2.32% and 3.49% in RMSEr over the GF-2 and GF-3 models, respectively. On the DSM dataset, the proposed model yielded the highest accuracy of AGB (RMSE = 25.69 t/ha, RMSEr = 16.53%). The DSM data were identified as the most important variable, due to mitigating the saturation effect observed in the optical and SAR images for a dense AGB estimation of the mangroves. The resulting map, derived from the most accurate model, was consistent with the results of field investigations and the mangrove plantation sequences. Our results indicated that the AGB can be accurately measured by integrating images from the optical, SAR, and DSM datasets to adequately represent canopy-related information, forest structures, and tree heights.

resolve the saturation problem [37][38][39][40]. Aerial photographs using the structure-from-motion (SfM) algorithm and the Unmanned Aerial Vehicle (UAV) platform are a low-cost option to measure the heights of trees [41]. The UAV-based digital surface models (DSM) can determine the relative tree height of mangrove forests, because they mainly grow over even terrain [42]. Hence, the integration of optical, SAR and UAV-based DSM data to represent the spectrum of the canopy, structure, and height of mangroves can improve the accuracy of the estimation of AGB for dense mangrove forests. Navarro et al. (2019) estimated mangrove AGB by combining UAV-based tree height, Sentinel-1, and Sentinel-2 images, and provided accurate estimates for young and sparse mangrove plantations [17]. However, the effectiveness of this approach needs to be examined further for dense and complicated mangroves.
This research aims to address the aforementioned gaps in the estimation of the AGB. We integrate images from multiple sources-GF2 optical, GF3 SAR, and fixed-wing UAV-based DSM data-to estimate the AGB of mangrove plantations. The objectives of this study are (1) to develop prediction models for the AGB using original and composite bands generated from optical, SAR, and UAV-based DSM data; (2) to evaluate the effectiveness of the AGB models and select the best one; (3) to determine the importance of the chosen parameters; and (4) to map the AGB to observe the spatial pattern of the biomass of a mangrove plantation in comparison with field surveys and its sequence of growth.

Study Area
This study was conducted in the mangrove nature reserve area of Guangdong Province in Dawei Bay of Qi'ao Island, in the Pearl River estuary of China (E 113°36′40″～113°39′15″，N 22°23′40″ ~ 22°27′38″) [43,44]. It is a typical tropical and subtropical wetland ecosystem. The area of the reserve is 5103 ha, covering about 700 ha of mangrove forests: the largest contiguous area of artificially planted mangroves in China [6,45]. Mangrove forests are characterized by high spatial variability that represents a dynamic landscape. The mangrove forests planted in the study area were composed mainly of a fast-growing species, Sonneratia apetala (S. apetala), which was introduced from Bengal. They belong to the woody mangrove species, with features of high adaptability and seed production [46]. The heights of their trees usually increase by about 1.5 m during each of the first few years [47,48]. They were the tallest tree species in the study area, ranging from 2 m to 20 m, and the diameter at breast height (DBH) of older trees can be as large as 30 cm.
S. apetala has been artificially planted on the island since 1999 to control invasive species, (i.e., Spartina alterniflora) and reconstruct mangrove forests. S. apetala generally has an afforestation specification of 1-2 m × 1.5 m with high densities. The tree ages range from 1 year to 17 years with high biomass variability and complication. Its afforestation process runs seaward from land, which implies a gradient distribution of tree age and AGB. The mangrove plantations had extended throughout the study area, and S. apetala had become the dominant species, covering more than 80% of the mangrove forest.

Field Investigation
We conducted field investigations to obtain biomass samples from June to July in 2016. To ensure the availability of samples to build estimation models for the AGB, we collected 127 samples along almost all the accessible tidal creeks to account for the variation in the biomass of all stages of tree growth. The samples were located in the low-, middle-, and high-intertidal zones (Figure 1d). For each sample, we measured the height and diameter at breast height (DBH) of each tree and recorded data for trees within a 10×10 m quadrat. Tree height was measured by a handheld laser range finder (precision of 1 m; Trueyard SP-1500H, Trueyard Optical Instruments co.), and DBH higher than 5 cm was recorded by a breast diameter ruler at 1.3 meters above the ground. To match the locations of the samples and overlaid images coordinates, the four vertices and center of each quadrat were recorded by a submeter-accurate GPS. The precise locations of the plots were recorded with the assistance of the GPS and high-resolution images (UAV images with 0.12m spatial resolution). Details of the locations of the quadrats, such as distance to the shore, were identified in and marked on the images. Using the measured tree heights and DBHs during the field survey, we calculated the AGB of each mangrove tree using allometric equations [49], and computed the sum of the AGB of all trees within a quadrat, to represent the AGB of a sample.

GF2 Optical Data
Gaofen-2 (GF-2) captures high-resolution images. It was launched by China National Space Administration (CNSA), Beijing, China, in August 2014. It has been applied to land monitoring, urban planning, and resource surveys [50]. GF-2 images have a panchromatic band (1-m resolution) and four multispectral bands (4-m resolution): red (R), green (G), blue (B), and near-infrared (NIR). We obtained GF-2 multispectral images on 15 February 2017 from Land Observation Satellite Data Service Platform (Figure 1a). The pre-processing of the GF-2 images-including radiation calibration, atmospheric correction, and geometric correction-was carried out using the ENVI 5.4.1 software package. Atmospheric correction of the images was carried out using the fast line-of-sight atmospheric analysis of the spectral hypercubes (FLAASH) model with the ENVI module. The images were also geo-rectified to a 1:10,000 topographic map using ground control points, to ensure that the position error was smaller than 0.5 pixels.
After preprocessing, the images were used to calculate four vegetation indices (VIs)-difference vegetation index (DVI), ratio vegetation index (RVI), normalized difference vegetation index (NDVI) and soil-adjusted vegetation index (SAVI)-as input variables to estimate the AGB of the mangrove forest (Table 1).
The Gaofen-3 (GF-3), launched in August 2016, was developed by the China National Space Administration (CNSA), and is the first Chinese satellite that collects multi-polarized C-band SAR data. GF-3 images are the only radar images in the Chinese High-resolution Earth Observation System [55]. They have 12 imaging modes, ranging from single to dual and full polarization, with a resolution of 1 to 500 m and have a revisiting period of 3.5 days at most to the same point on Earth. Such characteristics render the GF-3 suitable for resource monitoring.  Table 2). Based on the original data, the image-related data were preprocessed to preserve phase and amplitude information in the complex images. FP SAR images contain speckle noise and geometric distortions that can have a significant negative impact on features of the polarization of the target objects. Thus, a preprocessing sequence consisting of data import, multi-look processing, adaptive filtering, radiometric calibration, and the geometric correction was applied to the GF-3 images using the SARscape 5.4.1 module embedded into the ENVI software. The steps of the preprocessing are as follows: a) The metadata of the SAR images was imported to obtain the slant-range resolution and angle of incidence. The ground-range resolution was then calculated using the following: where Rground represents ground-range resolution (4.5480 m), Rslant is the slant-range resolution (2.2484 m), and θ is the incidence angle (29.6281°).
b) The multi-look method was applied to de-speckle and re-sample the full-polarization GF-3 images. The GF-3 SAR images had a 1 × 1 multilook (azimuth × range), and were resampled to a regular grid with 5×5 m pixels in terms of azimuth and range resolution (5.30 × 2.25 m). c) After the multi-look processing, the images were preprocessed using adaptive filters to reduce speckle and to enhance the edges and other features. We applied the refined Lee method of adaptive filtering by setting a 5×5 m filter window. d) Radiometric calibration was carried out to convert the intensity values to the calibrated backscattering coefficient σ o (dB) of the normalized radar using the following equation [56]: where DN represents the pixel values of the complex images, and K is the calibration constant. This equation referred to ALOS satellite processing. The radiometric calibration is still being explored to better process GF-3 SAR images. Previous studies used equation 2 to perform radiometric calibration, and demonstrated its utility and obtained satisfactory results [57,58]. Further study for radiometric calibration of Gaofen-3 images is needed. e) Finally, the corrected backscatter map was generated from the backscattering coefficients to reduce the negative effect of the incidence angle on the radar's backscatter. Geometric correction is then performed to match the positions of the ground control points selected in the GF2 images to corresponding points in a 1:10000 topographic map, with the positional error smaller than 0.5 pixels.
(2) Variables derived from GF-3 images The preprocessed images were used to acquire HV/HH-, VH/HH-, HV/VV-, and VH/VVpolarized data, by calculating the ratio of the backscattering coefficients of different polarimetric channels.
The FP SAR data allow us to identify the scattering mechanisms of different types that can significantly improve the depiction of features of the target object. This was achieved by polarimetric decomposition techniques, to separate the received signals of the radar. Such analysis can help repose a simpler object susceptible to an easier physical interpretation as a combination of the scattering [59]. Coherent decomposition is used to measure the scattering matrix by the responses of coherent scatters [60]. The targets of coherent scatter are analyzed based on the Sinclair matrix (S) representing all polarimetric information. With linear horizontal (H) and vertical (V) polarizations, the Sinclair matrix can be expressed as follows: The Pauli and Krogager decomposition approaches were used to analyze the targets of coherent scatter based on the Sinclair matrix [61]. Both could be applied to a homogeneous distribution of mangrove species in the study area [62]. Pauli decomposition was used to extract features of the polarization of the objects by defining different polarization fundamental matrices representing various types of objects. Pauli's polarimetric parameters were then decomposed into three elementary scattering mechanisms: odd-bounce scattering (P1), even-bounce scattering (P2), and volume scattering (P3). The Krogager decomposition aims to decompose the scattering matrix of a complex symmetric radar target into the physical interpretation of three components: sphere (KS1), diplane (KD3), and helix (KH2) [63].
The radar vegetation index (SAR-RVI), derived from FP SAR images, was used as a measure of the randomness of scattering from vegetation. It models the vegetation canopy as a collection of randomly oriented dipoles [64], and has yielded a good correlation between the SAR-RVI and the AGB. It was calculated by preprocessed FP data as follows [65]: where σHH, σHV, σVH, and σVV represent the backscattering coefficients of the polarimetric channels HH, HV, VH, and VV, respectively. In this study, the 11 predictors derived from GF-3 SAR images-the four backscattering coefficients (HH, HV, VH, and VV) of the FP channels, three Pauli polarimetric parameters (P1, P2, and P3), three Krogager polarimetric parameters (KS1, KD3, and KH2), and the SAR-RVI-were used as input parameters to build and predict the AGB of mangrove forests. They can reflect the different properties of mangrove forests. The backscatter of HH is linked to both trunk and crown biomass, HV and VH return crown biomass, VV is dominated by branch biomass [66,67]. For the ratio of backscattering, they can potentially reduce topographic effects and forest structural effects, thereby increasing estimation accuracy [68]. The SAR-RVI reflects the canopy vegetation characteristics [64,69]. The Pauli decomposition can be used to separate the scattering matrix into simpler scattering responses related to single bouncing (e.g., canopy surface), double bouncing (e.g., trunk) and volume scattering (e.g., crown) [70,71]. The Krogager decomposition is related to surface, two, and threesided corner reflectors [72]. The partial variables are shown in Figure 2.

UAV-Based DSM
DSM data were derived from a fixed-wing UAV with an onboard SONY NEX-5T camera and GPS/inertial measurement unit in 2016. The configuration of the UAV was set to an altitude of 400 m, 80% frontal overlap, and 60% side overlap. A total of 349 valid photographs were captured with geolocation and altitude embedded into the EXIF data. They were processed by the Agisoft PhotoScan Professional (64 bit) software. The DSM data were generated by overlapping photographs and the SfM photogrammetry algorithm and were exported at a resolution of 0.12 m (Figure 1c). A geometric correction was then executed by a 1:10,000 topographic map and ground control points.

Mangrove Classification Based on GF-2 and DSM Data
In this study, we just focus on the AGB estimation of mangrove plantation species, S. apetala. Their spatial distribution needed to be identified to map and predict AGB. The mangrove plantation (S. apetala), with a homogeneous distribution, was identified prior to AGB estimation. The plantation in the study area had distinctive traits, such as the tallest trees, from those of other mangrove species. The GF-2 and DSM data were integrated to extract characteristics of the mangrove plantation accurately. We collected over 700 samples for classification by field investigation. Half of them were used to build the classification models and the other half to validate them.
The multispectral and panchromatic bands of the GF-2 data were first fused by the pansharpening method. After image fusion, image objects were generated using a multi-resolution segmentation algorithm in eCognition Developer 9.0 software package. For each object, the mean values of the four bands and DSMs, Vis (DVI, RVI, NDVI, SAVI), and texture features (homogeneity, contrast, entropy, mean, and correlation) were calculated and used as input features. The random forest (RF) algorithm was used to train and build the classifier, using the input features and measured training samples. Finally, the RF models were used to predict maps of mangrove plantation.
Compared with the measured test samples, the overall accuracy of the RF models for mangrove species classification was 86.14%.

Modeling and Accuracy Assessment of AGB Estimation
Models for the estimation of the biomass of the mangrove forest were developed using the random forest regression algorithm (RFR), which is an ensemble machine learning technique that consists of a large number (ntree) of decision trees grown by bootstraps of the original samples [73]. Each node of decision trees is separated by a random subset of input variables (mtry). The final results of the prediction are obtained by averaging the individual predictions of all regression trees [74]. The importance of all input variables was measured by out-of-bag (OOB) samples using the RF model and quantified by mean decrease in accuracy (MDA) [75]. Each variable's MDA was calculated by the difference in OOB error between the original dataset and the dataset with randomly permutated variables. To reduce the randomness of the RF models, the mean importance values of the input variables were measured 50 times.
To integrate data from multiple sources, all variables derived from GF-2, GF-3, and DSM were resampled at a resolution of 4 m to correspond to the GF-2 images. The variables were used as input variables, and the measured AGB samples were used as output variables to build the RF models. The spatial distribution of the AGB across the study area was predicted and mapped by the built RF models. We employed iterated five-fold cross-validation by partitioning the AGB samples into five datasets, four of which were used for training and one for validation to ensure the stability, reliability, and generalization capability of the models. All five datasets were generated using stratified random sampling, which led them to represent the entire range of biomass values. The accuracies of the built models were assessed by the root-mean-square error (RMSE) and relative RMSE (RMSEr) calculated from the observed and predicted values of the AGB.
To qualify the effect of the input variables on the accuracy of estimation of the AGB, four experiments were conducted. RF models were built to this end by combining different types of variables. In experiment 1 (Expt. 1 for short), the model used eight variables derived from optical images of the GF-2, including the four bands, DVI, RVI, NDVI, and SAVI. In experiment 2 (Expt. 2), the model employed 15 variables derived from GF-3 SAR images: the four full polarizations (HH, HV, VV, and VH); the ratio of backscattering coefficients of different polarimetric channels (HV/HH, VH/HH, HV/VV, and VH/VV); Pauli decomposition (P1, P2, and P3); Krogager decomposition (KS1, KD3, and KH2); and the SAR-RVI. In experiment 3 (Expt. 3), the model used 23 variables through a combination of GF-2 optical and GF-3 SAR data. In experiment 4 (Expt. 4), the model used 24 variables by integrating GF2 optical, GF3 SAR, and UAV-based DSM data.

Workflow for Analyses
This study focuses on developing effective models of the AGB of a mangrove plantation based on images from multiple sources, including GF-2 optical, GF-3 SAR, and UAV-based DSM images. The models were built using a machine learning approach (i.e., random forest (RF)), and the input variables were derived from multiple datasets. The corresponding accuracies were examined to study the effect of the input variables on the monitoring of the AGB. Finally, the model with the highest accuracy was used to predict and map the spatial distribution of the AGB of mangrove plantations. The workflow is provided in Figure 3.

AGB from Field Sampling
Field data of the mangrove plantation were collected a few times in the study area, and a total of 127 sampling units were obtained. The AGB was calculated by the allometric equation of the specific species. The plantation S. apetala had a density of 1623 trees per ha. The heights of the tree ranged from 2 to 20 m, with an average of 13.64 m. As it is a fast-growing species, the AGB of the mangrove plantation ranged from 90.65 to 237.74 t/ha, with an average of 159.70 t/ha, exhibiting a wide extent (43.88 t/ha of standard deviations), owing to their different ages. The field data revealed decreasing trends of AGB values in accordance with the sequence of growth from shore to sea.

Importance of Input Variables for AGB Estimation
The importance of the input variables was quantified by the RF algorithm to evaluate the relationship between them and the AGB (Figure 4). The results showed that the most important variable was the UAV-based DSM, implying that it is key to the AGB estimation. The next mostimportant variables were the VIs (RVI, NDVI, etc.) derived from GF-2 optical images, followed by those (KD3, HV, etc.) derived from the GF-3 SAR images.

Results and Accuracy Assessment of AGB Model
The RF models were developed using the observed AGB as output variables and the variables derived from images from multiple sources as input variables. The RMSE and RMSEr were acquired by the observed and the predicted AGB values based on five-fold cross-validation. As shown in Table  3, the mean of the predicted AGB values in the four experiments was 156 t/ha, in line with the mean observed value. The model using input variables derived from the GF-3 SAR images yielded the lowest estimation accuracy of the AGB of mangroves, followed by the model that used the GF-2 optical images. The estimation accuracy of the model obtained through the integration of GF-2 and GF-3 images was better than the model that used either GF-2 or GF-3 data, with a reduction of 2.32% and 3.49% in RMSEr, respectively. The combination of GF-2, GF-3, and DSM data produced the highest accuracy (RMSE = 25.69 t/ha, RMSEr = 16.53%) of all models. A two-sided t-test revealed significant differences (p ≥ 0.95) in the predicted AGB values between models, using the combination of GF-2, GF-3, and DSM data, and those using only GF-2 or GF-3 images, but no significant difference (p < 0.95) was observed between the values obtained by models formed by a combination of GF-2, GF-3, and DSM data, and those obtained by models formed through the combination of GF-2 and GF-3 images.  16.53 Scatterplots of the predicted versus the measured AGB values are presented to show the accuracy of the models with different input variables using the RF algorithm and five-fold crossvalidation in the study area. As shown in Figure 5, the predicted AGB values of all models were above the 1:1 line at lower values, indicating that AGB values of the mangrove plantation had been overestimated, but they were the opposite at higher values. The coefficient of determination (R 2 ) of the model derived by integrating GF-2, GF-3, and UAV-based DSM data was 0.61, followed by the model obtained by a combination of GF-2 and GF-3 images. The other two models yielded lower R 2 values.

Mapping AGB of Mangrove Plantation
The model derived using a combination of GF-2, GF-3, and UAV-based DSM data produced the highest accuracy, and was used to map the spatial distribution of the AGB across the study area ( Figure 6). The AGB map exhibited significant spatial variability, ranging from 106.163 t/ha to 266.162 t/ha, with an average of 137.89 t/ha. The AGB values of the species of mangrove in the southeast of the study area were higher than those in the west and northwest. The biomass decreased progressively from shore to sea, and trees near the shore had larger canopies and higher AGB values because they were older than those in other areas. Trees growing outside the forest edge and off the shore were younger and exhibited smaller biomass. The mangrove plantation map was consistent with the results of the field surveys, visual interpretation of remotely sensed images, and prior knowledge of Qi'ao Island.

Discussion
Estimating the AGB of forests based on satellite remote-sensing images remains challenging for tropical and subtropical mangrove forests, owing to various factors that interfere in the relationship between the AGB and the variables of images, such as the complex nature of their environments and complex forest structures [17]. The AGB of a forest has a close relationship, with its canopy-related information, structure, and height of trees, where single remote-sensing data cannot simultaneously provide this information [76]. Previous studies have investigated the combination of optical images (Landsat, SPOT, Sentinel-2, etc.) and SAR images (Sentinel-1, ALOS, etc.) to improve the accuracy of estimation of the AGB of forests [17,18]. The feasibility and applicability of data from the new GF-2 and GF-3 satellites from the Chinese civilian High-definition Earth Observation Satellite (HDEOS) program, launched in 2014 and 2016, respectively, by the China National Space Administration (CNSA), need to be tested.

Overall Performance of Random Forest Model
The objective of this study was to develop an accurate and robust biomass estimation of mangrove plantations. A random forest model was selected to establish non-linear relationships between the AGB and the input variables, because of its ease of use and prediction accuracy [43]. Previous studies already used machine-learning algorithms, such as support vector regression (SVR) or artificial neural network (ANN) for AGB estimation, which produces satisfactory results.  investigated the applicability of RF, SVR, and ANN for remotely estimating wheat biomass, and the results indicated that the RF model produced more accurate estimates of wheat biomass than the SVR and ANN models at each stage [77].
A major advantage of RF is bootstrap sampling and variable sampling, in which the subset of all variables is randomly selected using the best split for each node of the standard regression tree. In these situations, The RF model can decrease the algorithm's risk for overfitting and multicollinearity, due to relative insensitivity to variations in input variables, thereby improving generalization and robustness to predicted data. Therefore, some variables in this study are correlated. However, as demonstrated by Cutler et al., the RF model is not sensitive to collinearity, and has the ability to model complex and nonlinear interactions among predictor variables [78]. This is helpful, as it is commonly hard to determine which variables need to be removed when two or more variables correlate with each other [79]. Therefore, the RF algorithm provides a useful exploratory and predictive tool for estimating mangrove biomass.

Contribution of Input Variables to Measuring AGB of Mangrove Plantation
This study addressed the above issues using GF-2 optical and GF-3 data. The results indicated that the potential of the optical images and C-band FP SAR images for the AGB prediction of artificially planted mangroves were similar. The SAR-based results have been slightly weaker than the results with optical data, mainly due to weaker spatial resolution of the available SAR data in this study (5×5 m pixel resolution after multi-look method processing). On the other hand, the radiometric calibration referred to ALOS satellite processing using the SARscape, which may cause the errors of biomass estimation [57,58]. The advantages of SAR images are their multi-temporal acquisitions and independence of cloud cover, making the atmospheric correction of optical images more difficult [18].
The model of the AGB developed using a combination of GF-2 and GF-3 images yielded a higher estimation accuracy than those built using GF-2 or GF-3 images alone. This finding is consistent with previous studies, which have noted that integrating optical and SAR images can improve the accuracy of estimation of the AGB of forests, mainly because factors influencing biomass estimation, such as canopy-related information (canopy density, vegetation status) and forest structure, can be reflected by them [18,80].
However, optical or SAR images often incur saturation problems in canopies owing to dense vegetation, leading to the underestimation of biomass [81]. When considering variables of the DSM based on GF-2 and GF-3 images, the accuracy of the AGB model improved by 2.7% in terms of RMSEr. DSM data were also identified as the most important variable by the RF algorithm, because they were collected on a similar date with field measurements and can reflect the relative height of trees in a mangrove plantation, which is important for biomass estimation [42]. Tree height is usually computed from corresponding digital terrain models (DTM) subtracted from digital surface models (DSM). However, the DTM for dense mangrove forests is unavailable, due to the inability to penetrate their dense and complex canopy structures. The DTM is a stable constant for mangrove forests, because they mainly grow over even terrain [42]. The DSM can thus be considered to measure the relative heights of mangrove trees instead of the canopy height model (CHM). Previous studies have shown that the DSM derived from SfM and aerial photographs can solve the saturation problem and improve the estimation accuracy of biomass [82,83].
In this study, the P3, B, G, and R variables are least important (Figure 4). These variables with relatively low importance may be caused by the insensitivity to AGB estimation or be affected by multicollinearity. For example, the near-infrared (NIR) band of optical images is more widely used to estimate vegetation biomass content, because of its spectral reflection features in green vegetation, and visible (red, green, and blue) bands are usually used to emphasize vegetation health and classification, thus, they may be insensitive to AGB estimation. However, if the features have a correlation, it can be challenging to rank the importance of the features.

Spatial Distribution Patterns of AGB of Mangrove Plantation
As an initial effort for restoration and reforestation, the mangrove species S. apetala was introduced to the study area in 1999. As a result of its ability to spread, the artificial planting of S. apetala is becoming increasingly controversial as it may invade other mangrove ecosystems. The map of the AGB of the mangrove plantation derived here can provide baseline data for subsequent analyses and applications (e.g., carbon sequestration). The AGB values were predicted and mapped by a model that used GF-2 optical, GF-3 SAR, and UAV-based DSM data. The S. apetala afforestation process runs seaward from land, which implies a gradient distribution of AGB. The spatial distribution of the resulting AGB corresponded to the sequence of the mangrove plantation over time.
To further verify the reliability of the model and understand the spatial distribution of the AGB, areas occupied by the mangrove plantation that had grown before 2011 were extracted by WorldView-2 images [43]. They exhibited significantly higher AGB values than the other areas, due to their age and fast growth. This is shown in Figure 7. The results thus indicate that the map of the AGB of the unevenly aged mangrove plantation showed a greater heterogeneity of AGB values.

Limitation and Sources of Errors
We found a reasonable relationship among GF-2 optical, GF-3 SAR images and, field measured AGB by using the RF algorithm. The combination of multi-source datasets (GF-2, GF-3 images, and DSM) yielded a higher estimation accuracy. However, there are some limitations and sources of errors from the AGB estimation using multi-source images, caused by position errors of geometric calibration, and the time difference of images and field measurements.
The errors of geometric calibration among multi-source images were unavoidable, though we used a geometric calibration to a 1:10,000 topographic map using ground control. The mismatch of the pixels derived from multi-source may cause the uncertainty of AGB estimation. Similarly, we got the AGB samples by field investigation, and the closest remotely sensed data available. The differences in time acquisitions may bring additional sources of errors in the AGB estimation, due to the fast-growing characteristic of the S. apetala. The study from Ren et al. (2010) suggested that AGB accumulation rates at the S. apetala plantations decreased with the stand ages [48]. The AGB accumulation rates were 20.3 t/ha from 4 to 5 years stand, 5.6 t/ha from 5 to 8 years stand, and 2.85 t/ha from 8 to 10 years stand, respectively. In our study area, most of the plantations have reached over 5 years stand [84], and previous study has demonstrated mean AGB accumulation of S. apetala over the study area to be 4.17 t/ha per year [85]. The average of observed values of this study is 155.4 t/ha, which may cause a 2.68% error of AGB-which is within the acceptable range-due to the different date between the field data and the images. Therefore, the inevitable difference of one year between the remote sensing images and in situ measurements can be deemed acceptable.

Conclusions
This study explored the potential of GF-2 optical, GF-3 SAR, and UAV-based DSM data for estimating the AGB of the mangrove plantation of Qi'ao Island in China. The AGB model generated using a combination of GF-2 and GF-3 images from the Chinese civilian HDEOS program yielded a higher accuracy than those of models using only one of these datasets, with a reduction of 2.32% and 3.49%, respectively, in RMSEr. When considering variables of the DSM derived from the UAV platform, the AGB model achieved the highest accuracy with a further reduction of 2.7% decrease in RMSEr. The DSM was the most important input variable for AGB estimation as it deals with saturation problems in optical and SAR images. The resulting AGB map agreed well with field surveys and the growing sequence of mangrove plantations. The results showed that accurate AGB models and spatial distribution maps of mangrove plantation can be obtained using the RF model, and images from multiple sources (GF-2 optical, GF-2 SAR, and UAV-based DSM data). The combination of these data provided canopy-related information, forest structures, and tree heights for AGB modeling.
The study focused on the integration of GF2 optical, GF3 SAR, and UAV data for estimating aboveground biomass in China's largest artificially planted mangroves. The methodology can be used to produce accurate AGB models of mangrove forests, which can be difficult to obtain by field investigation. The AGB maps of S. apetala can help measure mangrove carbon sinks and provide baseline data for REDD+ programs, due to mangrove plantation. Future studies should further examine and improve AGB estimation uncertainty, such as accurate radiometric calibration and noise estimation of GF-3.

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