Matsala M. Dynamics of forest cover within Chornobyl Exclusion Zone

CEZ forests are unique landscapes being protected and experiencing quite a limited anthropogenic impact, while traditional human management is hitherto insignificant there. Yet, these forests were initially planted with schemes common in previous century. Literally these consequences of past forest management are shaping current CEZ forest dynamics. It is profoundly linked to natural disturbance regimes which are changing over time (more frequent and severe wildfires, pest attacks and forest disease outbreaks, excess tree mortality due to abiotic stress and droughts), and to climate change under specific environmental conditions (namely, radioactive contamination). In this investigation the different machine learning methods were found as capable to predict the spatial distribution of forest biophysical parameters in CEZ (basal area and growing stock volume) with sufficient precision and accuracy. That is, root mean square error fluctuated between 49-71% of the mean for basal area and between 65-98% of the mean for growing stock volume. The smallest error was achieved by gradient boosting machines model. However, k-Nearest Neighbors method allowed to more adequately reflect an empirical distribution (according to cumulative distribution functions) and spatial semivariation of input data. Based on these results it is suggested that Sentinel-2 satellite images and a small training dataset (consisted of 102 temporary sample plots) are sufficient to examine the main differences between modeling approaches. Created maps of biometrical parameters of new forest stands occurred after disaster within abandoned farmlands can be used as a reference for the geospatial assessment of radionuclides deposited in tree stemwood. With 90Sr radionuclide as an example, it is shown that exactly the spatial resolution of soil contamination maps defines the precision of final product, which will be utilized for ecosystem services assessment in CEZ forests. Based on this research, data of dense time series of Landsat images spectrally processed with temporal segmentation algorithm LandTrendr seems to be a reliable geospatial information source for the remote monitoring of CEZ forests. More specifically, Random Forest classification model with a total 80% accuracy and 90% accuracy for the land cover class “forest” (based on the estimates for reference year 2017), still produced an accurate forest mask map having 89% accuracy on the validation 1988-year data. Disregarding the catastrophic wildfires happened in 1992, 2015-2016, and 2020, forest cover has increased by almost 1.5 times: from 41% in 1986 up to 59% in 2020. Classification model for the whole period of available satellite data is capable to capture forest cover loss in CEZ driven by natural disturbances and anthropogenic influence. That is, remote monitoring of forest cover changes based on developed model matched well with data obtained from global datasets such as Global Forest Change and HILDA. However, the vast majority of new forests occurred on abandoned farmlands was not reflected in these global products, which addresses to the additional advantage of developed approach while applying a remote monitoring of CEZ forest ecosystems.