Generalization of U-Net Semantic Segmentation for Forest Change Detection in South Korea Using Airborne Imagery

Abstract

Forest change detection is essential to prevent the secondary damage occurring by landslides causing profound results to the environment, ecosystem, and human society. The remote sensing technique is a solid candidate for identifying the spatial distribution of the forest. Even though the acquiring and processing of remote sensing images are costly and time- and labor-consuming, the development of open source data platforms relieved these burdens by providing free imagery. The open source images also accelerate the generation of algorithms with large datasets. Thus, this study evaluated the generalizability of forest change detection by using open source airborne images and the U-Net model. U-Net model is convolutional deep learning architecture to effectively extract the image features for semantic segmentation tasks. The airborne and tree annotation images of the capital area in South Korea were processed for building U-Net input, while the pre-trained U-Net structure was adopted and fine-tuned for model training. The U-Net model provided robust results of the segmentation that classified forest and non-forest regions, having pixel accuracies, F1 score, and intersection of union (IoU) of 0.99, 0.97, and 0.95, respectively. The optimal epoch and excluded ambiguous label contributed to maintaining virtuous segmentation of the forest region. In addition, this model could correct the false label images because of showing exact classification results when the training labels were incorrect. After that, by using the open map service, the well-trained U-Net model classified forest change regions of Chungcheong from 2009 to 2016, Gangwon from 2010 to 2019, Jeolla from 2008 to 2013, Gyeongsang from 2017 to 2019, and Jeju Island from 2008 to 2013. That is, the U-Net was capable of forest change detection in various regions of South Korea at different times, despite the training on the model with only the images of the capital area. Overall, this study demonstrated the generalizability of a deep learning model for accurate forest change detection.

1. Introduction

Forest change has occurred through natural disasters or anthropogenic activities. The prolonged deforestation causes serious consequences for the environment, ecosystem, and socio-economic aspects [1,2,3]. In particular, the forest change region coupled with climate change is vulnerable to landslides [4,5,6]; this introduced the critical roles of the forest monitoring that included the improvement in the knowledge for forest function, health, and resilience at various spatial and temporal scales and for practical application for the forest modelling simulation. Therefore, forest monitoring is essential to identify the potential landslide hazard areas for mitigating the additional damage.

Remote sensing is the effective monitoring method for forest mapping, using optical, microwave, and thermal sensors in aerial and satellite platforms [7]. This promising technique provides the detailed spatial and temporal distribution of the forest on a consistent basis. Especially, the use of the optical spectra is a typical way for forest monitoring and management [8]. Onishi and Ise (2021) [9] utilized Red–Green–Blue (RGB) bands of unmanned aerial vehicle images for mapping the trees in order to manage the forest. Mäyrä et al. (2021) [10] applied the RGB images of airborne hyperspectral image to classify tree species in order to implement the forest survey. Adjognon et al. (2019) [11] introduced tree cover mapping using sentinel-2 satellite imagery with RGB bands to monitor the forest. For utilizing the remote sensing imagery to detect the forest change, image processing such as radiometric calibration, atmospheric correction, and geometric correction and image annotation such as labelling tree species are required. These synoptic monitoring and image processing are considered costly, time consuming, and labor intensive processes.

However, the open source data platforms have been developed in order to mitigate this limitation by freely providing the remote sensing images for enhancing derivative research. Vu et al. (2021) [12] emphasized the open-sourced data availability for urban functional land use mapping by only using free medium-resolution satellite images. Long et al. (2021) [13] discussed the significance of the image dataset construction framework for remote sensing image application and interpretation. Open source data platform accelerates the related studies, applications, and generalization by providing robust datasets. In particular, AI-Hub provides high resolution airborne images and the corresponding annotation images for machine learning studies and applications. The numerous airborne data released the expense and monitoring burden. In addition, the pre-preprocessing and labeling procedure by the expert institute guaranteed the quality of both airborne and labeling imagery. The reliability of the retrieved datasets is essential to conduct data-driven studies with significant performance.

Deep learning is a prospective technique that provides robust performance when dealing with complex datasets [14]. Convolutional neural networks (CNNs) have been introduced as representative deep learning models that provide reliable performance for feature extraction from multi-dimensional data [15,16]. In particular, the convolutional layer of CNN is able to directly process 3-dimensional satellite or airborne images without further procedure such as vectorization, which results in high computational burden due to the massive amount of the trainable weights. In addition, the internal kernels of convolutional layer compose the shared weights and slide the image inputs to efficiently extract the spatial and spectral features of multi- and hyper-spectral imagery. Khan et al. (2017) [17] concluded that CNN optimally extracted satellite image features. Sun et al. (2022) [18] showed effective spatial and spectral features, learning of a CNN-based approach by processing aircraft imagery. Based on this rich hierarchical feature extraction framework, the CNN model enables remarkable progress in pixel-wise semantic segmentation. The semantic segmentation is to assign a categorical label to every pixel in an image, which plays an important role in forest image classification. In particular, U-Net was introduced to provide a superb semantic segmentation performance using a fully convolutional network and dense skip architecture. The U-Net model was applied for the segmentation of biomedical images [19,20,21], self-driving images [22,23,24], and remote sensing images [25,26,27]. In relation to the forest environment, Wagner et al. (2019) [28] utilized U-Net segmentation to classify the forest types and tree species of regional scales. In addition, Zhang et al. (2021)[29] performed semantic segmentation of U-Net for forest fire detection. With open science trending globally, Zhong et al. (2020)[30] adopted open source data fusion with high resolution satellite images and Google Earth images in order to conduct semantic object segmentation of urban land use. Chen et al. (2021) [31] also implemented semantic segmentation of urban land use by using open big dataset such as OpenStreetMap and sentinel images. Hermosilla et al. (2022) [32] successfully applied the free Landsat archive to perform land cover mapping of wide range and long term. Moreover, the open source datasets contribute to improving the generalization of the model. Otálora et al. (2019) [33] emphasized the necessity of the large public datasets for evaluating the generalization of the deep learning algorithm. Vetrò et al. (2016) [34] also introduced open government datasets enhancing the long-term generalizability of the results. However, generalization of U-Net forest change detection with open source data has rarely been investigated.

Therefore, motivated by the robust semantic segmentation of U-Net architecture, this study assessed the generalization of forest detection with both the U-Net model and open source image datasets. The specific procedures for accomplishing the object were: (1) to pre-process the open source airborne images and forest label images in the capital area of South Korea; (2) to fine-tune the pre-trained U-Net structure for segmenting forest and non-forest regions; and (3) to test the forest change detection test of the generalization ability of the trained U-Net model basis, using open map service.

2. Materials and Methods

2.1. Overview

The overall research framework of this study is presented in Figure 1. To accomplish the forest change detection, the open source airborne images were collected with RGB format while the corresponding forest label images were acquired by having forest species annotation in each pixel. Both airborne and label images were then segmented to the specific size. In addition, the forest labels were categorized into two classes including forest region and non-forest region. Herein, the label of undistinguished regions was excluded. After that, the pre-processed imagery was fed into the deep learning model that was U-Net architecture for performing semantic segmentation that classifies forest and non-forest regions in each image pixel. This architecture contained a pre-trained module part, while another part was fine-tuned for the successful model training. The training, validation, and test performances of U-Net determined the capability of the forest change detection. The generalizability test images were prepared from the open map service platform and processed by the aforementioned procedure. Thus, this study evaluated the generalization of forest change detection using a deep learning approach with multi-dimensional images from open source data platforms. The following section describes the relevant information for each step in detail.

2.2. Study Area

The capital area of South Korea includes Seoul city, Incheon, and Gyeonggi province. This is one of the highest urbanized areas, where 51% of the entire population lives. Seoul has forest areas of 15,323 ha with a forest rate of 25%. Incheon and Gyeonggi show forest areas of 39,373 and 512,105 ha by having forest ratios of 37% and 50%, respectively. That is, the urbanization development is inevitable for maintaining the sustainability of the capital area. In the meantime, the proportion of forest region can be decreased, which would potentially result in an adverse impact on the wellness of the citizens and urban ecosystem such as heat island [35]. Furthermore, the urbanization of the capital area leads to the increase in the areas that are vulnerable to the natural disasters such as typhoon, intensive rainfall, and abnormal drying due to the global climate change. That is, forest regions can be damaged by landslides and forest fires. In Seoul, the Mt. Umyeon landslide destroyed forest and residential areas [36]. In the Gyeonggi area, an average of 74.4 forest fires occurred during the recent 10 years from 2010 to 2019 [37].

2.3. Airborne and Tree Annotation Data Acquisition

This study utilized the open data source platform for the change detection of forest regions using a deep learning model. AI-Hub (https://www.aihub.or.kr accessed on 4 October 2021) was developed by the National Information Agency (NIA). This platform has provided the datasets for machine learning research, applications, and education in various fields including vision, language, land environment, livestock, fisheries, safety, self-driving, education, and healthcare. The airborne imagery capturing the forest region from 2018 to 2019 can be found in the land environment datasets. The airborne data with the size of 512 × 512 have spatial resolution of 0.25 m and RGB format in the capital area. Moreover, AI-Hub provides the label images the same with the location, resolution, and numbers of the airborne images. The labels contain the forest tree species such as pine tree, larch, conifers, broadleaf trees, and the other conifers. Moreover, non-forest regions and indistinguishable areas are included in the labels. Thus, 1000 images with respect to the airborne and label images were used for the deep learning training, and 100 images of airborne and label were utilized for the validation. The image data quality was managed and controlled by the National Geographic Information Institute (NGII) of South Korea, including radiometric calibration and geometric correction. The labeling image data quality was managed by the Korea Forestry Promotion Institute (KOFPI).

2.4. Image Data Preprocessing

The selected airborne and label images were cropped to the size of 128 × 128 for feeding to the deep learning model, because this study adopted pre-trained deep learning model architecture that allowed for the aforementioned image size in terms of the input. That is, each image of 512 × 512 size produced 16 cropped images. The final numbers of training and validation data were 16,000 and 1600 images, respectively. Moreover, the labels of pine tree, larch, conifers, broadleaf trees, and the other conifers were unified as forest region denoting to 0. On the other hand, the label for non-forest regions was assigned to 1. If the indistinguishable areas were in the label data, both the airborne and label image data were excluded from the training and validation dataset because of avoiding the uncertainty of deep learning model training performance.

2.5. Deep Learning for Change Detection of Forest Region

This study applied a deep learning model for the semantic segmentation of the prepared input images to classify the forest region. The semantic segmentation allocates the certain label in each pixel of the image. Herein, U-Net architecture was adopted to perform the segmentation task classifying forest and non-forest labels in airborne imagery, to allow for the change detection in the woodland.

2.5.1. U-Net

U-Net was originally developed for the segmentation of biomedical images [21]. This model showed robust semantic segmentation performance to classify the human cells. The architecture of U-Net composes contracting and expanding paths. In this study, MobileNetV2 [38] was applied to the encoding path, while pix2pix [39] was used to the decoding path. MobileNetV2 has the strength of feature representation using thin bottleneck residual layers (Figure 2). Pix2pix framework allows learning from input image to output image mapping. These paths extract input features using fully convolutional layers that allow the feeding of arbitrary sizes for input images. In addition, the convolutional feature extraction has the advantage of extracting the contextual information of the input images by using an internal kernel operator with learnable weights and biases (Equation (1)).

p_i = f_a(O_i-1 × w_i + b_i)

(1)

where p_i is the output feature at ith convolutional layer, O_i-1 is output feature at i-1 layer, f_a is activation function, w_i is learnable weight at ith layer, and b_i is bias at ith layer.

The downsampling path learns the contexts of airborne images and the upsampling path then performs localization, which can provide a class label into each pixel. In particular, the skip connection adds the contextual features from each encoding layer to the equivalent decoding layer to mitigate context loss during the localization.

The airborne images and the corresponding label images were fed into the U-Net model for the training with the stochastic gradient decent method [40]. With pixel-wise softmax function, the train loss is calculated by cross entropy function (Equations (2) and (3)).

$${\mathrm{s}}_{\mathrm{i}}\left(\mathrm{x}\right)=\frac{{\mathrm{e}}^{{\mathrm{p}}_{\mathrm{i}}\left(\mathrm{x}\right)}}{{{\displaystyle \sum }}_{\mathrm{j}=1}^{\mathrm{I}}{\mathrm{e}}^{{\mathrm{p}}_{\mathrm{j}}\left(\mathrm{x}\right)}}$$

(2)

$$\mathrm{C}={\displaystyle \sum }_{\mathrm{x}\in \mathsf{\Omega }}\log \left({\mathrm{s}}_{\mathrm{k}\left(\mathrm{x}\right)}\left(\mathrm{x}\right)\right)$$

(3)

where ${\mathrm{s}}_{\mathrm{i}}\left(\mathrm{x}\right)$ is the approximated maximum function, ${\mathrm{p}}_{\mathrm{i}}\left(\mathrm{x}\right)$ is nonlinear convolutional features by activation in channel i at the pixel location x, I is number of labels, C is cross entropy at each position of ${\mathrm{s}}_{\mathrm{k}\left(\mathrm{x}\right)}\left(\mathrm{x}\right)$ deviation from 1, k is the true class of each pixel (i.e., Ω→{1, …, I}).

This study utilized the pre-trained MobileNetV2 for the contracting path while it fine-tuned the number of kernels in pix2pix layers for the expending path. The transfer learning allows for fast adaptation during new task learning based on starting from pre-assigned initial weight [41]. In addition, the fine-tuning of the baseline model contributes to improving its performance without making a complex internal structure [42,43]. Specifically, the numbers of internal kernels in expanding paths were changed from 512 to 32 at the first convolutional layer of pix2pix, 256 to 16 at the second layer, 128 to 8 at the third layer, and 64 to 4 at the fourth layer, respectively. Moreover, for the other hyperparameter of U-Net training, the momentum constant was 0.9 and the learning rate was assigned as 0.00001. Furthermore, the training batch size and epoch were set as 256 and 2000, respectively.

2.5.2. Forest Change Detection of U-Net

The trained U-Net model was applied to new images for detecting forest land changes in order to evaluate the model generalizability, and this study used spatial information open platforms such as Kakao map (https://map.kakao.com accessed on 4 October 2021). This web service can provide the satellite images of South Korea with various spatial resolutions and different time periods. These images are originally captured, managed, and offered by NGII. The regions containing woodland changes (except the capital area) were selected for testing the U-Net model (Figure 3). In Kakao map we randomly searched specific regions in Chungcheong, Gangwon, Jella, Gyeongsang, and Jeju Island, where the forest changes were observed by time (i.e., the image in 2009 and image in 2016 had different topology for forest and non-forest regions). We prepared these images with spatial resolution of 30 m by zooming in at the website, and then downloaded and cropped the images at 128 × 128 size. The prepared test data applied to the trained U-Net model. The segmented results were recombined to the original size of the image, after which the evaluation of U-Net change detection was conducted by the visual comparison between the segmentation maps and ground truth maps.

2.6. Performance Evaluation

Segmentation performance is evaluated by calculating pixel accuracy, F1 score, and intersection over union (IoU). These accuracy indexes indicate the correspondence between ground truth labels and modeled label for each pixel of the image. Thus, the pixel accuracy, F1 score, and IoU are defined by the equation as follows:

$$\mathrm{PA}=\frac{{\mathrm{TP}}_{\mathrm{n}}+{\mathrm{TN}}_{\mathrm{n}}}{{\mathrm{TP}}_{\mathrm{n}}+{\mathrm{TN}}_{\mathrm{n}}+{\mathrm{FP}}_{\mathrm{n}}+{\mathrm{FN}}_{\mathrm{n}}}$$

(4)

$$\mathrm{F}1\mathrm{score}=\frac{{\mathrm{TP}}_{\mathrm{n}}}{{\mathrm{TP}}_{\mathrm{n}}+\frac{1}{2}\left({\mathrm{FP}}_{\mathrm{n}}+{\mathrm{FN}}_{\mathrm{n}}\right)}$$

(5)

$$\mathrm{IoU}=\frac{{\mathrm{TP}}_{\mathrm{n}}}{{\mathrm{TP}}_{\mathrm{n}}+{\mathrm{FP}}_{\mathrm{n}}+{\mathrm{FN}}_{\mathrm{n}}}$$

(6)

where PA indicates pixel accuracy; ${\mathrm{TP}}_{\mathrm{n}}$ is the number of true positive that pixels are correctly labelled as forest region; ${\mathrm{TN}}_{\mathrm{n}}$ is the number of true negative that pixels are incorrectly labelled as forest; ${\mathrm{FP}}_{\mathrm{n}}$ is the number of false positive that pixels are correctly labelled as not forest; and ${\mathrm{FN}}_{\mathrm{n}}$ is the number of false negative that pixels are incorrectly labelled as not forest.

3. Results and Discussions

3.1. Open Data Source for the Semantic Segmentation of Forest Region

The acquired images contained the various topological information of the forest region. The tree species in the label data composed of pine tree, larch, broadleaf tree, needleleaf tree, and the other conifers. Among these trees, pine tree is the major tree species in South Korea. However, in Gyeonggi province the dominant tree species is broadleaf tree. This species is a representative tree in the central region of South Korea as natural vegetation [44]. In data processing of the tree label data, all of the tree species were categorized to one class as forest region, while the other areas were assigned to non-forest region. The selected datasets for the forest classification with U-Net showed a higher portion of forest class than the non-forest. Specifically, the training datasets composed of forest region and non-forest region of 88% and 12%, respectively. For the validation data, the ratios of forest area and non-forest area were observed as 89% and 11%. Although these data had relatively large amounts of the information of the forest compared to the trait quantity of the non-forest, the distinctive spectral features between forest and non-forest regions could work as indications of the feature extraction during the deep learning training.

3.2. Forest and Non-Forest Region Segmentation of U-Net

Semantic segmentation of the U-Net model classified the forest and non-forest regions in the capital areas. The training and validation losses were converged to 0.0014 and 0.27, respectively (Figure 4a). In addition, the pixel accuracies of the U-Net training and validation were 0.99 and 0.97 (Figure 4b). Specifically, in the training results, the visual analysis of U-Net segmentation images showed the identical forest and non-forest regions compared to RGB and the label imagery (Figure 5). From the visual comparison of the validation results, the U-Net model followed well the regional trends of the forest and non-forest (Figure 6). Zhang et al. (2021) [29] demonstrated that the U-Net was capable of segmenting wildfire progression using the combination of radar and optical images showing overall accuracy of 0.93. Ahmed et al. (2021) [45] introduced the U-Net accuracy of 0.95 for object segmentation in the drone image scenery. Hacıefendioğlu et al. (2022) [46] also found that the capacity of U-Net application for detecting ice from wind turbine blade images had the accuracy of 0.97. That is, the U-Net segmentation results from these previous studies imply that the U-Net model is an effective candidate for semantic segmentation tasks for various fields [47]. Moreover, the merits of U-Net for segmentation tasks were better per-pixel localization and global representation of the input features, by conducting the expending path, than the fully CNN models [48]. However, for the validation result, shadowing regions were misclassified as forest region. The similar features between the shadow and shadowed tree might make the feature extraction of U-Net confused (Figure 6c,i). When the shadowing effect was dominant, the shadow removal process was needed to improve the performance of semantic segmentation [49]. Moreover, grass areas were improperly classified to the regions of the forest due to the optical traits such as normalized difference in vegetation index (Figure 6f) [50].

This study selected the trained U-Net model at the epoch of 1500, although the loss curve of the validation was not reached at the minimum but kept increasing while the loss of the training was at the minimum. This loss trend was typically called because the training model had undergone to overfitting. Thus, we analyzed the U-Net model performance at the epoch of 500, 1000, 1500, and 2000. The validation pixel accuracies of the selected epoch were 0.96, 0.97, 0.97, and 0.97, respectively (

Table 1. Summary of U-Net performance of each epoch.

	Training		Validation
Epoch	Loss Value *	Accuracy	Loss Value	Accuracy
500	0.0263	0.99	0.13	0.96
1000	0.0063	0.99	0.18	0.97
1500	0.0020	0.99	0.24	0.97
2000	0.0014	0.99	0.27	0.97

* is unitless.

). The improvement in the validation performance and the visual comparison between the U-Net and label images implied that the results at the epoch of 1500 were solidified, so as not to be under the overfitting (Figure 6c,f,i). Thus, the overall result of the U-Net model at the epoch of 1500 showed robust classification performance of forest and non-forest regions with the pixel accuracy, F1 score, and IoU of 0.99, 0.97, and 0.95, respectively (

Table 2. Overall performance evaluation of the selected U-Net model.

	Accuracy	F1 Score	IoU
Semantic segmentation of forest and non-forest	0.99	0.97	0.95

). The authors of Dong et al. (2019) [51] presented the entire results of the semantic segmentation with pixel accuracy and F1 score values of 0.89 and 0.89, while Mi et al. (2020) [52] introduced maximum accuracy of the semantic segmentation by having an F1 score of 0.93 and IoU of 0.90.

In some cases we found that the trained U-Net model correctly classified the forest and non-forest regions, but the label data were incorrect (Figure 7). This might contribute to the increase in the validation loss. Li et al. (2018) [53] introduced the increasing trend of validation accuracy, despite reaching the global minima of training loss. Even though the designed deep learning architecture successfully provided the detection performance of glaucomatous optic neuropathy, this previous study suggested the larger datasets for mitigating the conflicting validation tendency. Moreover, when determining the optimal model selection, the overall results with respect to the number of epoch needed to be considered [54]. Lee et al. (2019) [55] adopted the deep learning model architecture that had relatively low training loss, but high validation and test loss compared to the other architecture showing the low validation and test loss.

3.3. U-Net Test Performance Evaluation for Forest Change Detection

The trained U-Net model was applied for the change detection of forest areas using open map service. In order to evaluate the model capability, the test areas were selected from the major provinces of South Korea including Chungcheong, Gangwon, Jeolla, and Gyeongsang provinces. The area in Jeju Island was also tested for the change detection. The test performance was evaluated by the visual comparison between the ground truth images (Figure 8a,c,e,g) and the model output images (Figure 8b,d,f,h). The U-Net model was able to detect the forest area change in Chungcheong from 2009 to 2016 that might be due to the urban development (Figure 9a). This model also detected the change in the forest regions in the middle of the mountainous sector in Gangwon from 2010 to 2019 (Figure 9b), Jeolla from 2008 to 2013 (Figure 9c), Gyeongsang provinces from 2017 to 2019 (Figure 9d), and where the decrease in the forest zone might be caused by anthropogenic development or natural disaster. In case of not being mainland, the change area of the forest in Jeju Island from 2008 to 2013 due to the expansion of the residential areas was able to be detected by the segmentation model (Figure 9e). Overall, the U-Net model showed the significant capacity to generally detect forest change regions (Figure 9). That is, the features of tree species in the capital area could cover the features in the other areas in the varied provinces and island. In fact, the tree species of South Korea mainly occupy broadleaf trees and pine trees as 47% and 22%, respectively [56]. Thus, for evaluating the generalizability of the U-Net model, the volunteering geographic information service, including Kakao map and Google map, mitigated further data production [30]. Choi et al. (2020) [57] utilized Kakao map images for training the U-Net resulting in a significant localization performance of urban building areas. Moreover, several previous studies observed satisfactory image segmentation accuracy by coupling Google map images with U-Net architectures [42,58,59]. The test performance was dependent on the training epoch. Even though the high validation loss was observed in the high number of epochs, the performance of the forest change detection was improved until the epoch of 1500 with respect to the test areas (Figure 10d).

Furthermore, the U-Net performance was affected by whether the indistinguishable label was excluded or not. The test performance of the U-Net model with the annotation of indistinguishable region was shown in Figure 11. The training and validation accuracies were 0.90 and 0.59, respectively. The features of the unknown area might confuse the overall feature extraction of the model training, thereby resulting in lower segmentation performance of U-Net than the performance without the unspecified label data. Ishibashi et al. (2021) [60] suggested that vague or superimposition labels should be excluded when building the input datasets in order to improve deep learning model performance. Hang et al. (2021) [61] conducted additional input image processing to rectify the incorrect label to prevent the ambiguous deep learning model training. Thus, the thorough data quality control is the challenge for minimizing the uncertainty of the deep learning model performance because the model is vulnerable to the anonymous label and annotation mistakes [62,63].

3.4. Feasibility Implication of Using Open Source Data and Deep Learning for Change Detection

This study identified the generalization ability of the U-Net model for forest change detection using open source airborne imagery and open map service. The open source datasets are based on the stability in that the data are securely managed [64], the continuity in that the data keep updating, and the reliability in that data quality is controlled [65]. Moreover, the open map service can maximize the accessibility of the data utilization with the varied aspects of space and time [66,67]. Additionally, the granularity of the open image data allowed elucidating the various aspects of the forest change [68]. Large numbers of datasets can amplify the applicability of a data-driven approach for the change detection [69]. These traits of open source data facilitate the generalized deep learning model for the map generalization process [70]. Moreover, change detection mapping based on the open source data provided effective means in terms of the identification of topological transformation in large areas [71]. Thus, the generalization ability of U-Net can be in accordance with reality due to the larger data quantity by covering broader spatial ranges [72].

In case of applying the deep learning model, the use of pre-trained architecture lowers barriers to deep learning application in the ways of lightweight [73,74]. Additionally, its fine-tuning approach enables the ability to improve the robustness of pre-trained deep neural networks [75]. The fine-tuning of the pre-trained model was actively applied to the change detection tasks including vehicle detection [76], damaged building detection [77], and lexical detection [78]. Therefore, open data and deep learning model usage contributed to enhancing the feasibility, not only for the change detection but also the other derivative studies. Furthermore, although this study demonstrated the robust feasibility of open source data and deep learning applications for forest change detection, the lag time of the image acquisition could miss important change circumstances. In this matter, further studies will be needed for the real time change detection of forest areas caused by floods, landslides, and earthquakes with acquiring additional instant image data from aerial platforms such as drones.

4. Conclusions

This study assessed the feasibility of open source airborne images and the U-Net model to examine the forest change detection in the regions of South Korea. The open source data mitigated the physical limitations of acquiring the remote sensing and contributed to improving the generalization of the data-driven approach, while the fine-tuning of the pre-trained U-Net model improved the deep learning model utility. The U-Net model showed good training and validation accuracy for classifying forest and non-forest regions at each image pixel. Herein, the training steps and the excluding ambiguous label were critical to semantic segmentation performance of U-Net in terms of forest regions. Moreover, the generalization of the trained U-Net model followed well the spatial and temporal trends of forest distribution changes in South Korea. This robustness of the U-Net change detection with few data implies that the training data should contain the representative information in terms of various tree species. Therefore, the U-Net feasibility in the open source data basis was proven to have a potential for forest change detection. Future research will be needed to conduct additional filtering of incorrect label data for minimizing the data uncertainty, and also to test extension of the forest change detection through deep learning models by mapping the vulnerable regions of forest change.