Classification of Deforested and Degraded Areas in North Korea Using Phenology-based Indices and Spatiotemporal Data Fusion Method

Abstract

Forest ecosystems provide ecological benefits to both humans and biodiversity. Long-term anthropogenic pressure on forests, and frequent disturbance such as forest fires, landslides, and droughts, combined with abiotic factors such as climate variability, create an unsustainable environment. North Korea, suffers from serious forest degradation and deforestation issues. Given that these issues are major threats to the present and future state of forest ecosystems, the country is home to some of the most degraded forests in the world. Poor agricultural practices in North Korea, such as the excessive use of pesticides and fertilizers, are damaging productive capacity of the land. Poor agricultural practices, deforestation, and overgrazing cause soil erosion and forest degradation. Climate change and extreme weather events impact North Koreas deforested lands and degraded forests considerably, resulting in serious soil loss, landslides, and damage to farmlands and agricultural production. Furthermore, adaptation inerventions (including infrastructure) to extreme climate events, are lacking, and healthy forest ecosystems are essential elements for sustainable environment in North Korea. To provide systemic, prioritized restoration and planning efforts to address these problems, it is essential to establish a classification or monitoring system for typical types of deforestation and degradation. Because land-cover maps for monitoring deforestation and degradation are among the most fundamental data used in many scientific fields, developing a cost-effective method for classifying forest cover with high accuracy and high spatial and temporal resolution where access to data is difficult or when data are not available, is still challenging and necessary for North Korea. For this purpose, a cost-effective method for classification and monitoring forest cover dynamics in high spatial and temporal resolutions was proposed in this study. First, to classify types of deforested and degraded areas and to increase accuracy, this study proposed an optimal combination of phenology-based multi-index distinctions, as well as ways to distinguish complex, heterogeneous land cover in forests (such as hillside fields and unstocked forests) from plateau vegetation and natural forests. The outcomes of this research extend beyond those of most previous studies, which have usually been focused only on dryland forest. Previous work also involved the use of single-image classifications based only on spectral data to distinguish types of deforestation, and consequently, had difficulty capturing the heterogeneous spectral signature of land-cover categories over large areas. The outcomes of this work can be summarized as follows. 1) The seasonal patterns indicated by three indices (Normalized Difference Vegetation Index, Normalized Difference Soil Index, and Normalized Difference Water Index or NDVI, NDSI, and NDWI, respectively) showed differences typical of each type of vegetative cover in forest land. Thus, it was possible to overcome the reflectance value confusion that occurs when using only one image and increase the classification accuracy. 2) To classify complex land cover and dynamics, Random Forest proved to be a useful tool for classifying a variety of input features. The results highlighted the types of deforested land and their distribution in North Korea. The classification result showed an overall accuracy of 89.38% when phenology-based indices were combined with Random Forest. 3) The phenology-based indices that resulted in classification greater than 20% are the NDSI during the growing season (from March to May), the NDVI during the end of the season (September), and the NDWI during the start and end of the season (March/April and October). Combining these variables can effectively classify or help monitor vegetative cover. Our method greatly improved accuracies for classification of heterogeneous vegetative cover and presented deforested areas more reliably. Therefore, it should be useful for continued monitoring of variation in forested areas during forest restoration efforts in North Korea. The ecological impacts of forest degradation in the study area should also be urgently considered. Second, to construct the continuous fine-resolution satellite data using the most accessible datasets, a spatiotemporal data-fusion method was developed to blend satellite images in heterogeneous and spectrum-changing areas. However, previous methods have shown difficulties in predicting spectrum-changed pixels in fine resolution, as each bands value changed in different ways during the period of input and prediction dates. To overcome this limitation, this study proposed a spectrum-correlation-based spatiotemporal data-fusion method (RDSFM), to blend temporally fine-resolution data with temporally dense coarse-resolution data. The RDSFM integrates ideas from unmixing-based methods and a homogeneous index in the FSDAF, IR-MAD, and weighted-function-based method into one framework. The RDSFM was tested using a real landscape and compared to the referred spatiotemporal method, namely FSDAF. The results of the accuracy assessment demonstrate that the RDSFM has higher accuracy, especially in fragmented areas and the NIR band, and the method also maintains more spatial details. The spectral change of each pixel solved in the RDSFM is more robust than that in the FSDAF because of the strategy of weights based on MAD. MAD addresses detection of nontrivial change in multi-bands and bi-temporal data based on canonical correlation analysis. To estimate the MAD-based weights, the MAD for detecting the temporal change used a fine-resolution image at t1 and a coarse-resolution image at t2. The MAD for detecting the changes occurred using different sensors, with a fine-resolution image at t1 and a coarse-resolution image at t2. This can effectively detect the relative alteration of each band in one coarse pixel. This method can help accurately predict the pixel value in the areas where the spectrum changes due to land cover or within-class variance. The RDSFM, like other spatiotemporal data-fusion methods, can be applied to analyze land-cover dynamics, monitor vegetation phenology, detect land-cover change, and identify where degradation has occurred. To determine the effectiveness of the aforementioned proposed methods, the most difficult regions for vegetative cover change monitoring were tested using a simple classifier-unsupervised classification. This part successfully demonstrated the techniques effectiveness and convenience. Using stacked phenology-based multi-variables with the RDSFM is a powerful means of reducing classification errors, and enables better characterization of complex land-cover change status at 30-m resolution. In this manner, the degradation and deforestation that occurred from 2001 to 2014 were detected in three cases: 1) degradation from forest to unstocked forest, 2) degradation from hillside farms to unstocked forest, and 3) deforestation from forest to hillside field. The classification result showed an overall accuracy of 86.1% when using the simple unsupervised classifier. It is clearly effective and convenient to perform annual land-cover change analysis to produce a transition matrix with no information. Finally, various applications for analyzing forest ecosystems were suggested, using information on land cover and detailed forest degradation in fine resolution. In summary, a combination of phenology-based indices with a spectrum-correlation-based spatiotemporal data-fusion method can improve the spatial and temporal resolution for classification or monitoring of deforestation and degradation. Furthermore, this study suggested that in monitoring forest cover dynamics, distinguishing forest degradation and deforestation is essential to systemic planning and various analyses of forest ecosystems.