DIGITALNA ARHIVA ŠUMARSKOG LISTA
prilagođeno pretraživanje po punom tekstu



ŠUMARSKI LIST 7-8/2023 str. 57     <-- 57 -->        PDF

forests in places where agricultural land has been abandoned. The abandonment of agricultural land in these regions is a consequence of social and political changes such as the disintegration of socialist agrarian policy and the joining of countries to the global market (Kuemmerle et al. 2008).
The use of remote sensing for assessment of forest resources is crucial because forestry engineers and forest institutions cannot rely on existing records from the real estate cadastre to estimate the size of forest resources. Remote sensing provides several benefits, including extensive coverage of the area, continuous and up-to-date availability of images, consistent outcomes that are not influenced by human bias, effortless incorporation with spatial data, cost-efficiency, and accessibility to a broad range of professionals (Lechner et al. 2020).
Machine learning has been used as a tool for processing remote sensing data. Machine learning can be seen as learning by example because it is an automatic approach to creating empirical models based only on data. Unlike the non-parametric approach, machine learning makes no assumptions about the data, its probability distribution, or its functional form. It is used to solve regression and classification problems (Kovačević 2021). The ones that are most often used are Decision Tree, Random Forest, Neural Network and Support-vector machine.
RF (Random Forest) is a technique that builds multiple decision trees, using a randomly selected subset of samples and training variables. RF is based on the idea that a combination of bootstrap aggregated classifiers perform better than a single classifier (Breiman 2001), where the bootstrap component means that each tree is parameterized by a series of randomly selected set of observations with replacement from the training data. Two parameters need to be set in order to produce the RF. The first parameter, Ntree, specifies the number of decision trees to be generated, while the second parameter, Mtry, determines the number of variables to be randomly selected and tested for the best split when growing the trees (Belgiu and Dragut 2016).
When compared to other methods, RF has a low computing complexity and is able to handle big data sets, process thousands of input variables, estimate the importance of a variable in the classification process, and be resistant to noise and limit values (Rodriguez-Galiano et al. 2012).
The RF machine learning technique is implemented in the CAST (Caret Applications for Spatio-Temporal) library for application within the programming language R. This library uses the CARET (Classification And REgression Training) functionalities which represent the most widely used package for model training and prediction using machine learning techniques (Kuhn 2019).
This article focuses on monitoring forest areas and the contradiction between official forest records and the situation on the ground, based on Sentinel-2 data. One cadastral municipality is tested to clarify the true reality of the forest resources in the Republic of Srpska entity. Additionally, to demonstrate how quickly changes take place, changes that happened between 2017 and 2022 in one cadastral municipality are displayed.