<abstract xmlns="http://www.w3.org/1999/xhtml">Landsat-8 OLI and Sentinel-2 MSI images from years 2015 and 2016, a 1:10,000 digital soil map and a large number of reference samples were used with a random forest machine learning implementation in GRASS GIS to construct a tree species map for the entire territory of Estonia (42,755 km2). Class probabilities for seven main tree species, an extra class for other species and probability of the forest cover not conforming to the forest definition were assigned for each pixel. Validation of dominant species distribution by area showed very strong correlation at county level both in state forests (R2 = 0.98) and in private forests (R2 = 0.93). Validation of tree species composition using harvester measurement data from 2,045 regeneration felling areas showed also very strong correlation (R2 = 0.75) with the measured values of the proportion of coniferous trees. There was some tendency to underestimate the proportion of more common species and overestimation was found for the species with smaller proportion in the mixture. The accuracy for the proportion of deciduous species that were present in a smaller number of reference observations was substantially smaller. Validation of the results by using data from 659 large sample plots from the database of the Estonian Network of Forest Research Plots and 3,002 small sample plots from the National Forest Inventory (NFI) data base confirmed the findings based on harvester data. The NFI data revealed also a decrease of estimation error with the increase of forest age. Cohen’s kappa index of agreement for main species for NFI sample plots with main species proportion equal to or greater than 75% decreased from 0.69 to 0.66 when observations with forests younger than 20 years were included in the comparison. Overall, the constructed map provides valuable data about tree species composition for the forests where no up to date inventory data are available or for the projects that require continuous cover of tree species data of known quality over the entire Estonia.</abstract>

Landsat-8 OLI and Sentinel-2 MSI images from years 2015 and 2016, a 1:10,000 digital soil map and a large number of reference samples were used with a random forest machine learning implementation in GRASS GIS to construct a tree species map for the entire territory of Estonia (42,755 km2). Class probabilities for seven main tree species, an extra class for other species and probability of the forest cover not conforming to the forest definition were assigned for each pixel. Validation of dominant species distribution by area showed very strong correlation at county level both in state forests (R2 = 0.98) and in private forests (R2 = 0.93). Validation of tree species composition using harvester measurement data from 2,045 regeneration felling areas showed also very strong correlation (R2 = 0.75) with the measured values of the proportion of coniferous trees. There was some tendency to underestimate the proportion of more common species and overestimation was found for the species with smaller proportion in the mixture. The accuracy for the proportion of deciduous species that were present in a smaller number of reference observations was substantially smaller. Validation of the results by using data from 659 large sample plots from the database of the Estonian Network of Forest Research Plots and 3,002 small sample plots from the National Forest Inventory (NFI) data base confirmed the findings based on harvester data. The NFI data revealed also a decrease of estimation error with the increase of forest age. Cohen’s kappa index of agreement for main species for NFI sample plots with main species proportion equal to or greater than 75% decreased from 0.69 to 0.66 when observations with forests younger than 20 years were included in the comparison. Overall, the constructed map provides valuable data about tree species composition for the forests where no up to date inventory data are available or for the projects that require continuous cover of tree species data of known quality over the entire Estonia.

Construction of tree species composition map of Estonia using multispectral satellite images, soil map and a random forest algorithm

Institute of Forestry and Rural Engineering

Forest Department, Estonian Environment Agency

Forestry Studies

This work is licensed under the Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License.

Landsat-8 OLI and Sentinel-2 MSI images from years 2015 and 2016, a 1:10,000 digital soil map and a large number of reference samples were used with a random...

Construction of tree species composition map of Estonia using multispectral satellite images, soil map and a random forest algorithm

Data publikacji: 25 sty 2019

Zakres stron: 5 - 24

Otrzymano: 06 mar 2018

Przyjęty: 30 lip 2018

DOI: https://doi.org/10.2478/fsmu-2018-0001

Słowa kluczowe
forest inventory, random forest, tree species, raster map

© 2018 Mait Lang et al., published by Sciendo

This work is licensed under the Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License.

Construction of tree species composition map of Estonia using multispectral satellite images, soil map and a random forest algorithm

Data publikacji: 25 sty 2019

Zakres stron: 5 - 24

Otrzymano: 06 mar 2018

Przyjęty: 30 lip 2018

DOI: https://doi.org/10.2478/fsmu-2018-0001

Słowa kluczoweforest inventory, random forest, tree species, raster map

© 2018 Mait Lang et al., published by Sciendo

This work is licensed under the Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License.

Słowa kluczowe
forest inventory, random forest, tree species, raster map