2.3.3. 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.

#### 2.3.4. 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 pan-sharpening 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%.

#### *2.4. 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 di fference 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 di fferent 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 coe fficients of di fferent 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.

#### *2.5. Workflow for Analyses*

This study focuses on developing e ffective 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 e ffect 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.

**Figure 3.** Workflow for the measurement of the aboveground biomass of artificially planted mangroves by integrating images from GF-2, GF-3, and UAV-based DSM datasets.
