Original scientific paper
https://doi.org/10.31298/sl.145.3-4.2
Deforestation monitoring with Sentinel 1 and Sentinel 2 images – the case study of Fruška gora (Serbia)
Dušan Jovanović
orcid.org/0000-0001-6834-0376
; dusanbuk@uns.ac.rs
Milan Gavrilović
orcid.org/0000-0003-4671-7544
; Univerzitet u Novom Sadu – Fakultet tehničkih nauka Novi Sad, Srbija
Mirko Borisov
orcid.org/0000-0002-7234-6372
; Univerzitet u Novom Sadu – Fakultet tehničkih nauka Novi Sad, Srbija
Miro Govedarica
orcid.org/0000-0003-1698-0800
; Univerzitet u Novom Sadu – Fakultet tehničkih nauka Novi Sad, Srbija
Abstract
Forest and forest ecosystems have a big importance for the whole living world on the earth. Rapid deforestation poses a great danger and increases the effects of climate change. Large forest areas are cut down every year around the world and these activities need to be closely monitored to reduce their negative impact. Knowledge of valid and current geospatial data on forests and forest areas, obtained by interpreting the data by remote sensing methods has great importance for rapid response and
management of forest areas. Decisions that are based on outdated and insufficiently precise data can have negative consequences.
The researched area of Fruška gora is located in Vojvodina and occupies the northern part of Srem. Due to its natural properties, it enjoys the status of a special nature reserve. Pastures and fertile land, vineyards and orchards, decorate the slopes and lower parts of Fruška gora, while the areas above 300 meters above sea level are covered with dense, deciduous forests.
This paper presents a method of analysis of radar Sentinel 1 SAR satellite images, together with a combination of multispectral Sentinel 2 images, with the aim of identifying missing and newly formed forest areas, as well as assessing the usability of free, for everyone available radar satellite images for forest observation.
The described methodology is based on the selection of areas of interest, the selection of radar images for the chosen time epoch, image processing, the selection of training sets by combining radar and multispectral images. The classification of radar images was performed on the Cloud platform using the Random Forest classification algorithm. The study showed that in each analysed period from 2016 to 2019, the area under missing forest is larger in relation to the newly created area under forests, as well as the growing trend of new forest areas.
Estimation of classification accuracy for each observed time epoch was performed by calculating the error matrix and Kappa statistics, and the average classification accuracy was about 97%. Visual analysis and comparison of the obtained results with historical data confirmed the high accuracy of identification of missing forest areas.
The presented method showed that RF classification of free Sentinel 1 and 2 satellite images, can be used as a reliable and up-to-date data for forest monitoring with satisfactory quality and very quickly.
Keywords
SAR; Copernicus; Random forest classification; forest monitoring; change detection
Hrčak ID:
257006
URI
Publication date:
30.4.2021.
Visits: 1.817 *