The purpose of this study was to improve the accuracy of forest cover classification in North Korea, which cannot be accessed by using random forest model. Indeed, another goal of this study was to analyze the change pattern of bare forest land in different ways. In addition, the distribution of forest-changing area into cropland, grassland and bare land was estimated for the bare forest land.
According to the study results, this method showed high accuracy in forest classification, and the method was also effective in analyzing the change detection of bare forest soil in North Korea for about 10 years.
Introduction
Background
In particular, there was intense damage to the Musan Rifle Range, Yonsa Rifle Range and Hoeryeoung City which was near the Tumen River in North Hamgyong Province. According to the Joint Assessment North Hamgyong Floods 2016 report of the United Nations Office for the Coordination of Humanitarian Affairs (UN OCHA), the victims were more than 50,000 families in Musan-gun area, and those of Yonsa-gun area and Hoeryong City was between 10,000 and 50,000 respectively (Figure 1).
Purpose of study
Like South Korea, North Korea is located in the monsoon climate zone, where precipitation is concentrated in the summer, so the damage in case of natural disasters caused by cleared forests is always greater. For example, severe flooding occurred in North Korea due to unusually heavy rains in September 2016.
Definition of denuded forest land
Literature Review
Lee (2004) reviewed the status of forest change and degraded forest area in North Korea by summarizing previously published materials. Especially, Johannes confirmed that the accuracy of forest degradation detection has been improved and the deforestation map has been prepared as a final product using the fused NDVI-PALSAR image. 2015) conducted a study to analyze land cover change in detail using the AViFS sensor, which is divided into color, tone, texture and context.
2011) conducted a study to analyze deforestation in the Amazon River basin using multi-sensor time series satellite data. Therefore, in the near future for a unified Korean peninsula, REDD will be a significant issue, and now it is significantly important to monitor the forest and deforestation status in North Korea. 2012) conducted a study to analyze a status of forest degradation in the area near Mount Geumgang in North Korea and to estimate carbon emissions therein by using Landsat TM. 2014) quantified forest loss in North Korea by estimating carbon emissions based on satellite remote sensing data.
Although many studies have been done so far on forest degradation by machine learning technique, there are not many related studies on North Korea. However, this study was a comprehensive analysis of the entire North Korea using Moderate Resolution Imaging Spectroradiometer (MODIS) low resolution satellite images, and it is necessary for analysis of more specific areas to improve the accuracy of classification. This study also attempted to increase the accuracy of land cover classification, including the mapping of bare forest soil using random forest machine learning approach using the Landsat array and high-resolution RapidEye satellite imagery.
Study Area and Study Data
Study area
It is a project to prevent forest loss while securing carbon credits by applying opportunity costs to the land that maintains its good quality forest area without losing or destroying it. Besides, machine learning approach has been steadily used in land cover classification based on remote sensing (Gislason et al., 2006; Rodriguez-Galiano et al., 2012; Pal Mahesh, 2005). Several studies have shown that random forest shows good performance in forest cover classification (Markus et al., 2012; Baccini et al., 2012).
Therefore, this study chose Musan-gun as a test area, which has been a severely degraded area in North Korea, and analyzed bare forest land there. Despite the widespread and extensive occurrence of deforestation throughout North Korean territory, Musan-gun was chosen as the research object region of this study because it is one of the areas where deforestation is most severe in North Korea. Musan-gun is geographically located on the north side of the border with China through the Tumen River, and as part of the Baekmu Plateau on the northwestern side of the Hamgyeong Mountain Range together with the Baekdu Plateau, it is an inland area surrounded by steep mountains above 1,500 m above sea level (Figure 3).
More than 90% of this mountainous area remains uninhabited, and it is also one of the coldest areas in northern Hamgyeong Province. The average annual temperature of Musan-gun is 5.2 ˚C, the average January temperature is -13.4 ˚C, the average July temperature is 21.3 ˚C, and the frost period is more than 200 days a year. Although Musan-gun was one of the most forested areas in North Korea, it is also one of the most damaged areas due to reckless exploitation of forests.
Study data
- Satellite images
- Reference data
The second period was the test year and during this period Landsat 8 OLI and RapidEye satellite images were used. The reason for separating the base year and the test year is to compare how forest cover has changed and deteriorated between two periods. For this purpose, the base year was chosen, taking into account 10 years earlier than a test year to allow changes to be easily confirmed.
It has 8 spectral bands such as Red, Green, Blue, NIR, and captures data through these bands with a resolution of 30 m. However, these should initially be considered cloud and cloud shadow, because those obstacles can have a significant effect on satellite image quality when using the Landsat dataset. In general, just like South Korea, North Korea also has a humid and humid climate in summer and therefore, unfortunately, no Landsat 8 images were acquired during the summer season.
Because it is possible to obtain data from the Red Edge wavelength, RapidEye is effective in various areas such as agricultural crop monitoring and forest monitoring. Unfortunately, it is difficult to obtain measurement data from relevant government organizations or to conduct direct field research. Therefore, Google Earth can be considered the only and most reliable reference data in North Korea.
Methodology
- Image preprocessing
- Modeling process
- Input variables
- Machine learning approach
Incidentally, the RapidEye satellite images used in this study were orthorectified Level 3A products, so the radiometric correction was applied, and also corrected by DEM based on appropriate GCPs. The overall modeling process applied in this study can be categorized into several parts as Figure 4. For this, FLAASH atmospheric correction was applied to Landsat 5, Landsat 8 and RapideEye satellite images using ENVI 5.2 software.
In this process, the preprocessed reflectance images were merged into a single image for each study site through the merge tool in ArcGIS version 10.4.1 software. One is to extract the sample points for each class and the other is to extract the input variables to use the next machine learning. In this study, 6 representative classes were selected, and these 6 classes consisted of forest, grass, cropland, built-up, bare soil, and water.
Moreover, topographical variables such as DEM, slope and aspect were also applied in this study. In this study, three tests were conducted and compared based on the period. In this scheme, Landsat 8 derived variables such as NDVI and NDBI per selected date were applied, as well as DEM, slope and aspect at 30 m resolution.
Study Results
Phenology pattern analysis of NDVI and NDBI input variables
Accuracy analysis and importance of variables
On the other hand, the test year overall accuracy was 0.74 when only Landsat 8 OLI-derived variables were used. During the test year, a total of 35 input variables were used, including topographic variables such as DEM, slope, aspect and the variables derived from 4 days of Landsat 8 imagery. In this case, the variables derived from the satellite images taken on 28 May 2016 were highly significant.
Also, the overall accuracy was 0.73 using both Landsat 8 OLI and RapidEye derived variables as input. The total number of input variables in this case was 59, including topographic variables such as DEM, slope, aspect and the variables derived from both 4-day Landsat 8 images and 4-day RapidEye images. In this case, the variables derived from the satellite images taken on May 28, 2016 were also highly significant, and NDVI variables were also important.
Final classification map
In this study, three main forest types in North Korea were categorized into each class as deforested cropland to crop, undeveloped forestland to grass, and bare land to bare land, respectively. Compared to a base year, the classes classified as bare forest land increased slightly during the test year. To analyze the bare forest land, this study first compares the classification map extracted through the machine learning random forest model and analyzed the overall forest area.
The study investigated the status of bare forest land in North Korea between the base year and test year based on 10-year changes. Among the areas estimated as bare forest land, especially the changing areas from forest to cropland were high in Musan-gun. In particular, Musan-gun has been one of the most vulnerable areas in North Korea in that it has suffered rapid forest degradation.
However, due to the isolated political situation, few studies on the relevant data measurement or monitoring of bare forest land in North Korea have been conducted so far. To monitor bare forest land, the use of remote sensing satellite images is also the only and most effective way in the case of North Korea because field surveying is not available there. Therefore, this study aims to develop an improved technique to classify land cover and bare forest land in North Korea.
In addition, this study aims to find out the improved and effective techniques for surveying bare forest land in North Korea by using the forest change detection method. Segment-based land cover classification using texture information in degraded forest land in North Korea.
