Accuracy assessment of global tree cover maps in the lowland plains of Ukraine

Abstract

Introduction. Recent advances in remote sensing technologies have provided new opportunities for collecting reliable data on forests. Free access to large archives of multispectral satellite images have altered the methods of forest monitoring and quality of forest maps. During last decades, we faced a number global forest mapping initiatives. Some of them are of great interest to Ukraine.

Analysis of recent research and publication. The global forest datasets are products of thematic classification of satellite imagery of different spatial resolution. The maps that were produced using sensors of coarse spatial resolution (more than 250 m) cannot be used for regional estimates of forest cover of Ukraine. On the other hand, several global land cover products at spatial resolution of 25–30 m became available during a few recent years. At least two of them developed by Hansen et al. (2013) and Sexton et al. (2013) are global forest cover datasets.

The classification of forests in a form of continuous fields of percent tree cover is quite popular approach (Berberoglu, Donmez, Ozkan, & Sunar, 2008; Coulston et al., 2012). The advantages of the approach is that continuous fields allow representing the spatial structure of forest cover that has been formed due to complex of environmental factors and human activity. Continuous forest maps can be easily transformed into discrete or binary ones using specific threshold of percent tree cover. The accuracy of global percent tree cover maps has analyzed in many publications (Pengra, Long, Dahal, Stehman, & Loveland, 2015; Sexton et al., 2015; Ontikov, Shchepashchenko, Karminov, Dyurauer, & Martynenko, 2016). We showed that Global Forest Change (Hansen et al., 2013) dataset can be used for mapping forests in Ukraine using 40 % of tree cover as a threshold value. However, detailed analysis of Landsat Tree Cover Continuous Fields (Sexton et al., 2013) has not been yet performed for Ukraine.

Objective. The aim of the paper is to compare the accuracy of global forest cover products at 30 m spatial resolution for the territory of lowland plains of Ukraine.

Methods. Two available for free public access global forest cover maps generated at 30 m spatial resolution were evaluated in this study. The first dataset Global Forest Change (GFC) was developed by Hansen et al. (2013) from series of TOA reflectance Landsat ETM+ images of the year 2000. Seasonal mosaics were created using per-pixel compositing approach. That allowed to extract series of phonological metrics and use them to predict percent tree cover for each 30 x 30 m pixel. The GFC dataset also includes three additional layers representing forest cover change: loss - total forest losses since 2000, lossyear - yearly forest losses and gain - total forest gain.

The second dataset Landsat Tree Cover Continuous Fields (LTCCF) used in the study was generated by Sexton et al. (2013). It provides per-pixel assessment of percent tree cover for each leaf-on Landsat scene converted to to surface reflectance. The regression models were developed separately for each Landsat scene using 250 m MODIS vegetation continuous fields as training dataset. Afterwards models were applied to Landsat images at 30 m resolution.

To compere the products we used the reference dataset that included about 4700 sampling points randomly distributed within 21 regions of Ukraine. The sampling size and spatial distribution of random points were projected following recommendation by Olofsson et al. (2014). Collect Earth plugin for Google Earth was used during visual interpretation of sampling points. After, the map values for both dataset were extracted using reference dataset of sampling points.

Results and discussion. Compared to GFC, interpretation of LTCCF is much easier task. The accuracy of forest/not-forest maps depends on selection correct value of percent tree cover as threshold for classification. It was found that 40 % threshold selected in our previous studies for classification of GFC data corresponds to 25 % for LTCCF map. One of disadvantages of LTCCF dataset is that tree cover ranges from 0 to 80 %, however few observations collected within not-forested area have map values 0 %. As a result, we concluded that the map underestimates canopy cover for forest stands, but overestimates it for not-forested areas. The distributions of sampling points by percent canopy cover are bimodal for forested and not-forested areas. As a result,  significant level of commission and omission errors takes place during classification of both datasets, but their magnitude tends to be higher for LTCCF. Nevertheless, selection the tree cover value of 40 % as threshold for classification is more reasonable for GFC while 25 % - for LTCCF dataset.

The accuracy of both maps is not enough to predict tree cover of windbreaks. LTCCF demonstrated poor accuracy to distinguish windbreaks within agricultural field using selected threshold. Thus,  GFC dataset is more precise for the territory of lowland plains of Ukraine.

In addition, we analyzed the distribution of sampling points within loss and gain classes of GFC dataset. It was found, that user`s accuracy for loss layer as big as 70% while for gain layer – only 42 %. The dataset incorrectly represents map values if combination of signals was as follows: loss → gain. The analysis showed that GFC data is not accurate enough to estimate forest dynamics. At the same time, the errors of loss and gain layers do not have great influence to total accuracy of GFC maps so far as they comprise very small portion of area.