Mapping tree species composition of forest stands using Landsat seasonal mosaics and sample-based forest inventory

Abstract

Mapping tree species composition of forest stands using satellite imagery has great importance for assessment of forest recourses and their dynamics. The majority of modern publications have pointed on efficiency of tree species identification using time series of Earth observations in conjunction with sample-based forest inventory. We used time series of Landsat 8 OLI imagery to map tree species composition of forest stands, meanwhile regional-level sample-based inventory of forest in Sumy region (oblast) have been incorporated as source of reference data. Based on tree stems measurements on 333 circular sample plots, basal areas of eight forest forming tree species have been assessed (such as pine, oak, maple, lime, birch, ash, aspen and alder). The satellite images have been filtered for time frame of 2014–2016 and composed into cloud-free mosaics for three seasons: April-October, summer, autumn. After cloudiness were masked, for summer and autumn mosaics we applied maximum value composing approach using NDVI value, while for April-October mosaic we selected 1st, 3rd quartiles and median values of pixels in the specified time range. In total, 38 spectral (bands of red, near- and short-infrared spectra, NDVI and TCT bands) and non-spectral (geographical coordinates, elevation above see level, topographic position) features used for classification. Forest stands of different tree species composition as well as dominant tree species has been mapped using k nearest neighbor (k-NN) imputation approach. In this paper, we explored the effectiveness of two technics: 1) mapping distribution of tree species within Sumy oblast; and 2) predicting tree species composition of forest stands. We have used Google Earth Engine implementation of k-NN algorithm to predict basal area for eight major tree species in the region. Euclidean distance has been applied as a measure of closeness to the nearest neighbor. We have used k = 1 to search for neighbors in a space of spectral features of Landsat imagery and ancillary non-spectral information. As a result, continuous per-pixel distribution of basal areas of eight tree species has been predicted. Based on their contribution to total basal area, dominant tree species as well as stand composition at 30 × 30 m pixel resolution have been derived. It was found that forest stands composed by 3-5 tree species are common for Sumy oblast. At the same time, the lowest tree species diversity typically occurs in pine forests, while stands formed by oak, ash, maple and lime have a grater species richness. We believe these factors have significantly impacted accuracy of mapping in northern, central and southern parts of the study area, since more complex vertical structure of forest stands in southwest arises discrepancies in tree species identification. Based on predictive mapping, the distribution of major forest forming species of forests in Sumy oblast have been established. Using per-pixel estimates of basal areas and mapped dominant tree species, we have compared forest stands area with official forest inventory data. We have found some underestimation of remote sensing-based assessment of forested areas occupied by pine and oak stands. On the one hand we refer that to the map errors, on the other hands such inconsistencies may be result of classifying forests in operational forest management by desired species composition (not by their actual contribution to the total basal area). Nevertheless, derived results have pointed on necessity of satellite-based forest inventory in Ukraine which improve characterization of forest cover at landscape level through wall-to-wall mapping. We also concluded the effectiveness of tree species mapping using time series of Landsat satellite imagery and Google Earth Engine cloud platform.

Keywords

• classification
• k nearest neighbors
• cloud computing
• forest assessment
• monitoring
• sampling
• time series
• basal area