ABSTRACT

Following the nuclear disaster of 1986, forests have established throughout the abandoned agricultural landscapes within Chernobyl Exclusion Zone (ChEZ). However, they are yet to be monitored properly. Their biometrical parameters need a robust assessment considering climate change mitigation potential and wildfire-induced risks. To predict basal area (BA) and growing stock volume (GSV) of these forests using spatially explicit approach, we utilized Sentinel-2 satellite data and three types of machine learning models (k-Nearest Neighbors (k-NN), Random Forest (RF) and Gradient Boosting Machine (GBM)). Root mean square error among all models ranged between 5.2 m² ha⁻¹ (49% of the mean) and 7.2 m² ha⁻¹ (71% of the mean), derived for BA by the GBM and k-NN models, respectively. While total and mean estimates of forest attributes were quite similar within an entire ChEZ, GBM approach outperformed other methods by predicting GSV more precisely when compared to local reference data. At the same time, k-NN approach has shown better performance in terms of preserving the initial empirical distribution and semivariation patterns. We concluded that k-NN method should be used for the spatial predictions of forest attributes, however, with a specific focus given on the training data set quality and profound model validation.

摘要

1986年发生核灾难后，切尔诺贝利禁区（ChEZ）内所有废弃的农业景观都遍布了森林。但是，尚未对其进行适当的监视。考虑到减缓气候变化的潜力和野火引发的风险，需要对它们的生物特征参数进行有力的评估。为了使用空间显式方法预测这些森林的基础面积（BA）和生长种群数量（GSV），我们利用了Sentinel-2卫星数据和三种类型的机器学习模型（k-最近邻居（k -NN），随机森林（ RF）和梯度增强机（GBM））。所有模型的均方根误差在5.2 m ²  ha ^-1（平均值的49％）和7.2 m ²  ha之间^-1（平均值的71％），分别通过GBM模型和k -NN模型得出BA 。虽然在整个ChEZ中森林属性的总和平均估计非常相似，但是与本地参考数据相比，GBM方法通过更精确地预测GSV优于其他方法。同时，在保留初始经验分布和半变量模式方面，k -NN方法表现出更好的性能。我们得出的结论是，应将k -NN方法用于森林属性的空间预测，但是要特别关注训练数据集的质量和深刻的模型验证。