Forest Vertical Structure Mapping from Bi-Seasonal Sentinel-2 Images and UAV-Derived DSM Using Random Forest, Support Vector Machine, and XGBoost

Abstract

Forest vertical structure is vital for comprehending ecosystems and biodiversity, in addition to fundamental forest information. Currently, the forest vertical structure is predominantly assessed via an in-situ method, which is not only difficult to apply to inaccessible locations or large areas but also costly and requires substantial human resources. Therefore, mapping systems based on remote sensing data have been actively explored. Recently, research on analyzing and classifying images using machine learning techniques has been actively conducted and applied to map the vertical structure of forests accurately. In this study, Sentinel-2 and digital surface model images were obtained on two different dates separated by approximately one month, and the spectral index and tree height maps were generated separately. Furthermore, according to the acquisition time, the input data were separated into cases 1 and 2, which were then combined to generate case 3. Using these data, forest vetical structure mapping models based on random forest, support vector machine, and extreme gradient boost(XGBoost)were generated. Consequently, nine models were generated, with the XGBoost model in Case 3 performing the best, with an average precision of 0.99 and an F1 score of 0.91. We confirmed that generating a forest vertical structure mapping model utilizing bi-seasonal data and an appropriate model can result in an accuracy of 90% or higher.

1. Introduction

For humans, forests are not only economically viable but also important ecological resources (Prasad and Kant, 2003; Wu et al., 2021). Accordingly, managing and preserving these forests is of paramount importance, with the forest vertical structure serving as fundamental information for their proper management (Pascual et al., 2008; Hirschmugl et al., 2023). In general, the forest vertical structure refers to information in the form of stories that represent the vertical distribution of biological clusters, and thus it is crucial in determining the productivity and biodiversity of forests and has been extensively studied (Latham et al., 1998; Zimble et al., 2003; Cloude and Papathanassiou, 2008). Hence, studying forests, including forest vertical structure, is considered important (Sun et al., 2008; Bohn and Huth, 2017). From the 18th century to the present, the in-situ method has predominantly been used to determine the vertical structure of forests. The in-situ method allows the accurate determination of forest vertical structure; however, it requires considerable time, money, and manpower, employing in large areas is difficult (because of inaccessible areas), and continuous updating is virtually impossible (Liang et al., 2019).

With the recent development of remote sensing technology via data acquired through platforms such as satellites, aircraft, and unmanned aerial vehicles (UAV), studies have explored the possibility of solving the limitations caused by in-situ methods (Pause et al., 2016; Andres et al., 2018). Kimes et al. (2006) used light detection and ranging data and multiangle radiometers to estimate forest vertical structure, whereas Feng et al. (2016) used X-band interferometry synthetic aperture radar data to extract forest vertical structure parameters. In addition, research is being actively conducted to improve the accuracy of forest vertical structure mapping using machine learning techniques, which have recently been in the spotlight in the field of image classification, and remote sensing image acquisition (Lee et al., 2020a; Lee et al., 2020b).Moreover, albeit infrequent, recent studies also demonstrate a trend toward identifying forest characteristics through the consideration of seasonal dynamics within forest ecosystems (Lee et al., 2023). However, the majority of extant studies have not adequately accounted for the possibility of temporal variability in the vertical structure of trees. Moreover, even in cases where such variability is recognized, using UAV imagery, due to its prohibitive acquisition costs, presents challenges for continuous application.

Plants, including trees, that constitute the forest are sensitive to temperature changes, and the forest vertical structure may be altered accordingly; however, research concerning this remains insufficient (Potter et al., 2001; Tang and Dubayah, 2017). Satellites, for example, Sentinel-2, can acquire time series images free of cost that are multispectral and cover large areas; however, forest vertical structure has been rarely analyzed using bi-seasonal images (Jiang et al., 2022). In this study, bi-seasonal images are defined as two images obtained at separate seasons for the same location, it contains various types of information on forests owing to differences in the surrounding environments (temperature, atmospheric conditions, etc.). Reflecting this diversity, a more inclusive forest vertical structure mapping performance can be expected (Yu et al., 2021). Furthermore, data containing vertical information on forests, such as digital surface models (DSM) and digital terrain models (DTM), can be useful for forest vertical structure mapping (Lee et al., 2019; Park et al., 2021).

Consequently, the present study aimed to accurately map forest vertical structure using bi-seasonal images and machine learning techniques. Using bi-seasonal images, we induced the forest vertical structure mapping model to reflect seasonal elements. To demonstrate the superiority of bi-seasonal images, we separated the data into three cases after appropriate processing to maximize the use of the data’s features. Likewise, the three most commonly employed machine learning algorithms, namely random forest (RF), support vector machine (SVM), and extreme gradient boost (XGBoost), are utilized. RF integrates multiple decision trees to provide stable and precise predictions; SVM discerns optimal decision boundaries in high-dimensional space for modeling intricate relationships; XGBoost attains superior predictive performance and rapid learning through gradient boosting (Baek and Jung 2021; Kwon et al., 2019). To ascertain the optimal approach for constructing a forest vertical structure mapping model using bi-seasonal imagery, a comprehensive examination and comparison of these three algorithms is conducted.

2. Materials

2.1. Study Area and Data

The study area was a forest located in Geundeok-myeon, Samcheok-si, Gangwon-do, Korea, with the Taebaek Mountains to the west and the East Sea to the east. Fig. 1(a) shows the location of the study area, and Fig. 1(b) shows a forest vertical structure map of the study area determined precisely using a field survey. Fig. 1(b) shows that the study area consists of one- to four-storied forests, which are disproportionately distributed, with ratios of 0.36%, 70.52%, 17.30%, and 3.03% for one-, two-, three-, and four-storied forests, respectively.

Fig. 1. Study area: (a) location of the study area and (b) forest vertical structure map.

Subsequently, Fig. 2 presents the seasonal Sentinel-2 images of the study area, depicting (a) acquisition during the fall and (b) acquisition during the winter. Upon comparing the two images, the disparities stemming from seasonal transitions are visually ascertainable, encompassing alterations in both the extent and condition of the tree coverage. Particularly noteworthy is the delineated region marked with a white box, where seasonal discrepancies are notably pronounced. This area, classified as two-storied based on the forest vertical structure map derived from field surveys (Fig. 1b), requires correction through suitable processing and learning methods, given the visual alterations resulting from seasonal transitions.

Fig. 2. Sentinel-2 images of the study area: (a) acquired in fall and (b) acquired in winter.

Table 1 shows the vegetation species comprising the forest vertical structure of the study area, which follows the general classification criteria of temperate areas (Huete, 2012). In the study area, a one-storied forest consists of a canopy, and two-storied forests consist of grass or shrubs with a canopy. Two of the grass, shrub, and understory comprise three-storied forests with a canopy, and a four-storied forest includes all types of structures.

Table 1. Information of vegetation species composing forest vertical structure

The data used in this study were obtained as follows: First, to confirm the alteration in the forest due to season changes, we surveyed a UAV equipped with an RX02 camera on October 22, 2018, and November 29, 2018. We acquired optic images by shooting at a low elevation of approximately 200 m and then we created a 3-dimensional green band point cloud through an image matching technique using high-resolution optic images, and DSM with a spatial resolution of 20 cm was generated based on this. The reason why the data gathering range is established from late October to late November is because the season in the research region changes from autumn to winter, allowing changes in the forest vertical structure such as major temperature variations and falling leaves to be seen. Next, we acquired a 5 m resolution DTM from the National Geographic Information Institute (NGII). This DTM is numerical topographic data generated along the contour line and was resampled by bilinear interpolation to 20 cm, which is the same spatial resolution as the DSM.

Finally, we acquired the Sentinel-2 images on a date close to that on which the UAV images were obtained. For a more accurate comparison, we attempted to acquire and use Sentinel-2 images on the same date as the UAV acquisition date; however, Sentinel-2 has a 5-day revisit time and its availability is affected by weather conditions and cloud effects. Therefore, along with these factors, we investigated the temperature of the study area and selected the optimal Sentinel-2 images. We also resampled Sentinel-2 images taken on November 1, 2018, and November 26, 2018, by bilinear interpolation to a 20 cm resolution, similar to the UAV images. The specifications of the Sentinel-2 image are provided in Table 2.

Table 2. Specifications of Sentinel-2 image

NIR: near-infrared, SWIR: short-wave infrared.

3. Methods

Fig. 3 delineates that this study proceeded through the following three steps:(a) generating the input data required for training the forest vertical structure classification model, (b) training the model using the input data, and (c) evaluating the performance of the forest vertical structure mapping model.

Fig. 3. Data flow of research: (a) process of generating input data, (b) process of training the classification model with machine learning techniques, and (c) process of performance evaluation.

In step (a), initially, the data was partitioned into three cases to confirm the improvement when using bi-seasonal images: (1) DSM from October 22, 2018, and Sentinel-2 images from November 1, 2018 (fall);(2) DSM from November 29, 2018, and Sentinel-2 images from November 26, 2018 (winter); and (3) images from all dates. In addition, RF, SVM, and XGBoost were used to select machine-learning techniques appropriate for forest vertical structure mapping. Subsequently, spectral index maps were generated, and canopy height maps were obtained using Sentinel-2 images, DSMs, and the DTM. We used the green normalized difference vegetation index (GNDVI), normalized difference moisture index (NDMI), normalized difference red edge index (NDRE), normalized difference vegetation index (NDVI), renormalized difference vegetation index (RDVI), and structure insensitive pigment index (SIPI) asspectral indices and filtered the canopy height map using the median and standard deviation.

In step (b), the training and verification data were split based on the generated input data, and the RF, SVM, and XGBoost models were trained using the training data. Prior to dividing, we performed min-max scaling normalization and divided the training and verification data into 20% and 80% of the total input data, respectively. Finally, in step (c), the performances of the model, such as average precision (AP) and F1 score using verification data, were evaluated, and forest vertical structure prediction maps were generated. Nine results were derived using three cases and three models, and all results were quantitatively compared according to the generated evaluation indicators.

3.1. Generating Input Data

3.1.1. Spectral Index Maps

The spectral index is a method of obtaining the desired information through appropriate calculations using the fact that each spectral band has a different reflectance depending on the type of material in an optical image and is used to emphasize the spectral features contained in the image (Huete, 2012; Xue and Su, 2017). Furthermore, it reduces the spatial distortion or shadow effects in the image (Zhang et al., 2015; Valeriano et al., 2016). Through pixel-based band calculations, the GNDVI, NDMI, NDRE, NDVI, RDVI, and SIPI indices were generated from the red, green, blue, red edge, near-infrared (NIR), and short-wave infrared (SWIR) bands. The formulae for the six types of spectral indices used in this study are listed in Table 3.

Table 3. Information from the spectral indices used in this study

GNDVI determines changes in chlorophyll in the vegetation through calculations of NIR and green bands (Candiago et al., 2015; Hernández-Clemente et al., 2017). GNDVI was used because energy imbalances occur between multistoried forests, resulting in differences in chlorophyll presence rates, which would substantially affect the classification of forest vertical structure (Fleischer, 1935; Ellsworth and Reich, 1993). NDMI measures the water content of vegetation; the higher the value, the more moisture the vegetation contains. The SWIR band is used by the NDMI to represent values based on the ratio of near-infrared to short-wave infrared and is useful for specifying changes in forest vertical structure (Healey et al., 2006). NDVI is the most commonly used spectral index for analyzing vegetation and calculates the ratio of the NIR and red bands to determine the density and vitality of the vegetation (Carlson and Ripley, 1997; Delegido et al., 2013).

NDRE is similar to NDVI in that it represents the vitality of vegetation; however, NDRE is more accurate in determining the vitality of vegetation with high chlorophyll concentrations using a red edge band that penetrates a few layers of leaf cells deeper than the red band (Boiarskii and Hasegawa, 2019). The higher the values of both NDVI and NDRE, the greater the density and vitality of the vegetation (Lee et al., 2020b). The RDVI alleviates the impact of the soil background by renormalizing the NDVI, which has a saturation problem caused by forest closure (Fu et al., 2013; Li, 2019). Because the study area was almost entirely covered by vegetation, lowering the background of the soil using the RDVI was determined to yield more effective results. SIPI determines the ratio of carotenoid pigments in chlorophyll, allowing us to understand vegetation stress caused by differences in solar radiation (Penuelas et al., 1995). In the study area, where the forest vertical structure exists, the amount of insolation between stories was inevitably different; therefore, we used SIPI as input data. Overall, we assigned spectral indices that could represent the features of the vegetation itself and the forest vertical structure. All generated spectral index maps were normalized using min-max scaling.

3.1.2. Canopy Height Maps

Forest vertical structure is caused by differences in vegetation height; therefore, canopy height is an important parameter. The canopy height maps were generated by subtracting the DTM from DSMs. This is possible because the study area is only a forest of vegetation; therefore, if the DSM, which represents the height of the surface and the object above it, is subtracted from the DTM, which represents the height of the surface excluding the object, only the canopy height remains. The resulting canopy height maps were derived from a high-resolution DSM model (20 cm) and the surface was rough. Therefore, we used a moving kernel of 51 × 51 pixels to perform median and standard deviation filtering (Kwon et al., 2017). By referring to the forest vertical structure classification criteria, the kernel size was determined to be appropriate for the study area. The median filter was used to emphasize the height of the canopy, and the standard deviation filter was used to represent the deviation between the heights of the canopy. Finally, min-max scaling was used to normalize the filtered canopy height maps.

3.2. Training Classification Model with Machine Learning Techniques

In this study, we used machine learning techniques to generate forest vertical structure classification models. To perform image classification effectively, three techniques were used: RF, SVM, and XGBoost, which represent ensemble bagging, ensemble boosting, and kernel methods, respectively. The hyperparameters of the models trained using each technique were optimized using a randomized search cross-validation method, which can optimize various parameters by efficiently utilizing computing resources and time (Bergstra and Bengio, 2012). A detailed description of each technique is provided below.

3.2.1. Random Forest

RF is a machine learning technique that employs the ensemble bagging method to create multiple decision trees in a dataset via bagging and to select the majority of results by classifying data into each tree simultaneously (Belgiu and Drăguţ, 2016). The use of multiple decision trees reduced the risk of overfitting (Pal, 2015). When training with the RF, hyperparameters such as those that set the depth of the decision tree and those that set the number of decision trees can be optimized to generate an improved model. The following formula is used to generate predictions using the RF algorithm:

\(\begin{align}\widehat{Y}=\frac{1}{N} \sum_{i=1}^{N} Y_{i}\end{align}\)       (1)

Ŷ represents the predicted output, N is the number of decision trees in the RF, and? Y_i is the output of the ith decision tree.

3.2.2. Support Vector Machine

SVM identifies the decision boundary to optimally classify the data in a vector space. Because SVM is an algorithm that linearly divides through the inner Euclidean space, when data are dispersed in a high-dimensional manner, a separate processing step is necessary (Cortes and Vapnik, 1995; Noble, 2006). When data are dispersed in a high-dimensional form, a separate technique called the kernel trick is necessary because SVM is an algorithm that linearly separates through the inner space in Euclidean space. The radial basis function (RBF) kernel was employed to minimize the dimensions by calculating the similarity and proximity of discrete data (Amari and Wu, 1999). The SVM classifier using the RBF kernel has the following formula:

\(\begin{align}f(x)=\sum_{i=1}^{n_{s y}} \alpha_{i} y_{\mathrm{i}} K\left(x, x_{i}\right)+b\end{align}\)      (2)

f(x) represents the predicted class label for input x, n_sv denotes the number of support vectors, α_i represents the corresponding nonnegative Lagrange multiplier for the support vectors, y_i represents the corresponding class label (1 or –1) for the support vectors, K(x, x_i) denotes the value of the kernel function between input x and support vector x_i, and b represents the bias term. Furthermore, RBF kernels are commonly defined as follows:

\(\begin{align}K\left(x, x_{i}\right)=\exp \left(-\frac{\left|x-x_{i}\right|^{2}}{2 \sigma^{2}}\right)\end{align}\)       (3)

x and x_i denote the distance between the input data, σ represents the width (hyperparameter) of the RBF kernel. This value is determined during the training process of the classifier and controls the flexibility of the decision boundary.

3.2.3. eXtreme Gradient Boost

XGBoost is a technique for enabling parallel learning with gradient boost algorithms that use ensemble boosting (Chen and Guestrin, 2016).When producing several decision trees, ensemble boosting reflects weights based on errors from prior trees during the training of the next tree. Gradient boosting is an ensemble boosting strategy that uses gradient descent. Each time a tree is trained, it further reflects the residual model to seek the minimum value of the error and the best model. XGBoost is exceptional in terms of speed and performance, and since its introduction in 2014, it has displayed the best performance among machine learning techniques. The XGBoost algorithm’s formula is as follows:

\(\begin{align}\widehat{Y}=\sum_{k=1}^{K} f_{k}(x)\end{align}\)       (4)

Ŷ represents the predicted output, K is the number of weak models(decision trees) in the ensemble, and f_k(x) represents the prediction of the kth weak model for input x.

Table 4. Confusion matrix in this study

3.3. Performance Evaluation

The performance of the models generated using machine learning techniques was evaluated using precision-recall curves and F1 scores based on the precision and recall acquired via confusion matrix creation (Visa et al., 2011).

3.3.1. Precision-Recall Curve

The Precision-Recall Curve (PRC) is a graph with x-axis recall and y-axis precision that is extensively used to evaluate the performance of machine learning models (Buckland and Gey, 1994). Precision is the ratio of the model’s predicted true pixel to the entire predicted true pixel, and recall is the ratio of the model’s predicted true pixel to the entire actual true pixel. PRC is a graph that represents the precision and recall gained by shifting the threshold from 0 to 1, and both precision and recall are high, indicating that the model is better when the graph is closer to the top right. It can assess performance by calculating the bottom area of the graph, which is known as the average precision. The precision and recall calculation formulas are as follows:

\(\begin{align}\text {Precision}=\frac{\text { True Positive }}{\text { True Positive }+ \text { False Positive }}\end{align}\)       (5)

\(\begin{align}\text {Recall}=\frac{\text { True Positive }}{\text { True Positive }+ \text { False Negative }}\end{align}\)       (6)

3.3.2. F1 Score

The F1 score corrects the arithmetic error caused by the imbalance between the two indicators using a harmonized mean of precision and recall (Chicco and Jurman, 2020). It is frequently used to evaluate the classification accuracy of machine learning models because it provides more objective indications that are not influenced by precision or recall bias. The F1 score was calculated using the following formula:

\(\begin{align}\text {F1 Score}=2 \times\text {Precision} \times \frac{\text { Recall }}{\text { Precision }+ \text { Recall }}\end{align}\)       (7)

4. Result

4.1. Input Data

Machine learning is a technique that analyzes the characteristics and distribution of data with various properties to determine the best pattern. Eliminating the range and units of data is beneficial for model performance. Since, in particular, SVM is an algorithm that finds the boundaries of data in a high-dimensional space, normalization is required for optimal performance and speed. Consequently, all input data utilized in this study were normalized to values between 0 and 1 using min-max normalization. It is a normalization approach that simply modifies the distribution of data between 0-1 and can benefit from normalization while keeping the data’s properties. Furthermore, to avoid overfitting caused by data monotony, we randomly selected values while separating the training and test data to create datasets that were not limited to one part of the input dataset.

Fig. 4 shows the input data for Case 1 (October 22, 2018, and November 1, 2018). Figs. 4(a–f) shows the GNDVI, NDMI, NDVI, SIPI, NDRE, and RDVI maps created using Sentinel-2 images, and Figs. 4(g, h) shows a canopy height map with a median filter and a standard deviation filter, respectively. Fig. 5 shows the input data for Case 2 (November 26, 2018, and November 29, 2018). Figs. 5(a–f) shows the spectral index maps, and similar to Figs. 4(g, h), Figs. 5(g, h) shows the filtered canopy height map.

Fig. 4. Input data of Case 1: (a) GNDVI map, (b) NDMI map, (c) NDVI map, (d) SIPI map, (e) NDRE map, (f) RDVI map, (g) median filtered canopy height map, and (h) standard deviation filtered canopy height map.

Fig. 5. Input data of Case 2: (a) GNDVI map, (b) NDMI map, (c) NDVI map, (d) SIPI map, (e) NDRE map, (f) RDVI map, (g) median filtered canopy height map, and (h) standard deviation filtered canopy height map.

Because two stored forests were predominant in the study area, all data are in a similar format, but some maps can visually demonstrate the variation in forest vertical structure. Visual inspection revealed that the data demonstrating the difference in forest vertical structure in some regions are shown in Figs. 4(b, f, g) and Figs. 5(b, c, f, g). However, confirming the major difference in the other data was difficult. Furthermore, it can be observed that the areas where differences due to seasonal variations were visually evident when comparing the Sentinel-2 images of the study area have been significantly rectified, except for the image in Fig. 5(b). Fig. 5(b) represents the NDMI map of the study area obtained during winter using Sentinel-2 imagery, wherein a significant portion of the vegetation in the area has undergone changes. Given the relatively dry conditions during winter, characteristic of the NDMI spectral index used to estimate the health status of vegetation based on plant and soil moisture, it appears that the correction was somewhat less effective.

Moreover, compared to the data in Fig. 4, the data in Fig. 5 revealed a higher difference in values produced by variables other than the difference in forest vertical structure. This appears to be due to changes in forest conditions and leaf density caused by seasonal and temperature changes caused by differences in data acquisition dates.

4.2. Predicted Maps

As was previously stated, Case 1 used the data shown in Fig. 4, Case 2 used the data shown in Fig. 5, and Case 3 used both types of data. The hyperparameters were tuned so that the RF, SVM, and XGBoost models could grasp the properties of the data with varying characteristics. Randomized search cross-validation, which is a method of randomly extracting values within a specific range and deriving optimal hyperparameter values via cross-validation, was used to tune the hyperparameters. Each model has different types of hyperparameters and the number of hyperparameters varies. XGBoost can generate a more specific model by adjusting various hyperparameters. In this study, RF and SVM were tuned as two of the most frequent hyperparameters. XGBoost, on the other hand, tuned unique hyperparameters such as learning rate and min_sum_weight to take use of the ability to adjust various hyperparameters. Table 5 lists the hyperparameter tuning values for each model and case.

Table 5. Optimal hyperparameters for each model and case

The prediction results of the forest vertical structure classification model generated using prior techniques are shown in Fig. 6. Figs. 6(a–c) depicts the predictions of the RF, SVM, and XGBoost models for Case 1, Figs. 6(d–f) depicts the RF, SVM, and XGBoost models for Case 2; and Figs. 6(g–i) depict the RF, SVM, and XGBoost models for Case 3. Initially, the results of the visual analysis were as follows: 1) The prediction maps of RF and XGBoost using the identical ensemble technique revealed comparable tendencies, but XGBoost appeared to be more accurate, 2) although the SVM prediction maps were impressive, they appeared to be heavily influenced by noise or external elements of the forest vertical structure, and 3) the difference between Cases 1 and 2 was minor, but the prediction maps for Case 3 appeared to improve significantly.

Fig. 6. Forest vertical structure predicted maps: (a-c) RF, SVM, and XGBoost models in Case 1, respectively. (d-f) RF, SVM, and XGBoost models in Case 2, respectively. (g-i) RF, SVM, and XGBoost models in Case 3, respectively. Three white boxes (A, B, and C) indicate areas where the difference is most noticeable in the predicted results.

These results are highlighted in white boxes A, B, and C in Fig. 6. First, as shown in box A, despite the pronounced impact of seasonal variations, all predicted maps nearly precisely classified the vertical structure of the forest as two-storied. However, the SVM had more erroneously predicted values distributed in the two-story area and the area influenced by external factors in the input data than the other models. In box B, XGBoost exceeded RF in terms of accuracy, and SVM performed impressively. Finally, box C shows that the Case 3 prediction maps are greatly improved when compared to those of Cases 1 and 2.

4.3. Performance Evaluation

Fig. 7 shows the PRC generated to evaluate the performance of the nine models. Figs. 7(a–c), Figs. 7(d–f), and Figs. 7(g–i) represent the RF, SVM, and XGBoost PRC in Cases 1, 2, and 3, respectively. The yellow line of each PRC is a micro-average; the blue, orange, green, and red lines represent one, two, three, and four stories, respectively. Analyzing the graph’s shape and AP values yields an estimate of the overall performance and classspecific performance of an individual case and model (as seen in the lower left corner of the graph). Overall, compared to RF and XGBoost, the PRC of SVM was somewhat rigid. This implies that the model is unstable; that is, the classification performance is poor.

Fig. 7. Precision-recall curves: (a-c) RF, SVM, and XGBoost models in Case 1, respectively. (d-f) RF, SVM, and XGBoost models in Case 2, respectively. (g-i) RF, SVM, and XGBoost models in Case 3, respectively.

This was also corroborated by the AP of the lower area of the PRC; in all cases and classes, the AP of SVM was lower than that of RF and XGBoost. In addition, the PRC in Case 3 shows superior results in all models compared to the other cases, which also shows a slight difference between the AP in Cases 1 and 2, while the AP in Case 3 showed substantial improvement. In summary, SVM performed the worst of the models, RF and XGBoost performed similarly, and XGBoost performed better. When comparing each case, integrating the data from Cases 1 and 2 with different acquisition dates considerably enhanced the performance.

Table 6 lists the evaluation metrics derived from the confusion matrix for each case-by-case model. In terms of the F1 score, SVM outperformed RF, unlike the previously assessed PRC. This could be because RF is good at precision but lacks recall, whereas SVM is less precise than RF but has considerably better recall performance; therefore, the harmonic average of the two values, the F1 score, is high. In contrast, XGBoost outperformed RF and SVM in all cases, with the highest F1 score reaching 0.91, particularly in Case 3. In addition to XGBoost, both RF and SVM improved the F1 scores significantly in Case 3 when compared with Cases 1 and 2. In summary, 1) the best performance among RF, SVM, and XGBoost in the forest vertical structure mapping model was achieved by XGBoost, and 2) the performance was dramatically improved when data from different dates were integrated and utilized. Moreover, 3) the performance improvement shrank more when the input data were different than when the models were different, and 4) a mapping model with an AP of up to 0.9904 and an F1 score of 0.9101 could be implemented.

Table 6. Precision, Recall, and F1 score for each model and case

5. Discussion

This study’s main objective is to enhance the forest vertical structure mapping model by using bi-seasonal data that reflects plant seasonality. Additionally, using the three most popular machine learning models, it seeks to identify the optimum method for leveraging bi-seasonal data. Looking at the study’s findings, it becomes apparent that despite the utilization of data showcasing varying characteristics corresponding to seasonal changes, it was demonstrated that the utilization of bi-seasonal data (Case 3) overall improved the model’s quantitative assessment score. These results appear to have been achieved through appropriate data processing and model training. Moreover, we might draw the conclusion that using XGBoost is advisable for mapping forest vertical structure using bi-seasonal data because the XGBoost models performed the best generally. This is an impressive outcome given that bi-seasonal data has only seldom been used in previous research to map forest vertical structure. We also specifically proposed a strategy that may be employed entirely in future investigations, taking into account the utilization of freely available Sentinel-2 images.

Nevertheless, this research revealed the following constraints and uncertainties: 1) It is probable that factors other than seasonal variations have been at work since the shooting dates of the images obtained from UAV and Sentinel-2 are different. 2) Due to the relatively low resolution of Sentinel-2 images and the DTM resampling to a high resolution of 20 cm, resulting in pixels of high resolution, yet relatively low quality. 3) There is room for interpretation that Case 3’s improved performance was brought about by the utilization of extra data. To remedy this, advancements should be made in future research starting with the planning stage of data gathering. The second and third constraints can also be improved by utilizing techniques such as augmenting or acquiring extra data. Moreover, the second constraint could be further enhanced through the recent advancements in super-resolution techniques, allowing for a more refined resampling process. Future studies that reflect these developments will be more solid and convincing

6. Conclusions

This study investigated an optimal forest vertical structure mapping model based on machine learning algorithms using bi-seasonal remote sensing data. First, the study was divided into three cases to compare the results for each date with those obtained by integrating the data. The six bands were then utilized to build spectral index maps of GNDVI, NDMI, NDVI, SIPI, NDRE, and RDVI to leverage the spectral features of Sentinel-2, which offers images of various spectral bands. Furthermore, because the entire research region was forested, a canopy height map was obtained by subtracting the DSM and DTM, and a filtered canopy height map was generated by applying a median and standard deviation filter to emphasize the boundary.

We trained three types of machine learning models: RF, SVM, and XGBoost, and nine mapping models were derived. All models generated a confusion matrix to measure precision and recall, following which the PRC and F1 scores were compared to assess performance. When the PRC form and bottom area AP of each model were compared, the SVM revealed that the model was less stable, RF and XGBoost were similar, and XGBoost had a slight advantage.

Furthermore, using bi-seasonal data substantially improved the overall performance in Case 3 rather than highlighting differences. Similarly, when the F1 score was confirmed, RF performed worse than SVM, but XGBoost still performed the best. This was the same as in Case 3, which showed the best results when viewed casewise. Consequently, the best-performing mapping model was the XGBoost model in Case 3, with an AP of 0.9904 and an F1 score of 0.9101. If the forest vertical structure mapping model can continuously show classification accuracy of 0.9 or better through continuing study, the in-situ method of forest mapping could be replaced in the future. In addition, deep learning techniques, such as patch-based algorithms, must be employed to utilize the spatial characteristics of the data and perform more complex operations.