Fractional cover mapping of spruce and pine at 1 ha resolution combining very high and medium spatial resolution satellite imagery

Abstract

Increases in extreme weather events associated with climate change have the potential to put currently healthy forests at risk. One option to minimize this risk is the application of forest management measures aimed at generating species mixtures predicted to be more resilient to these threats. In order to apply such measures appropriately, forest managers need up-to-date, accurate and consistent forest maps at relatively fine spatial resolutions. Cost efficiency is a major factor when creating such maps. Taking European spruce (Picea abies) and Scots pine (Pinus sylvestris) as an example, this paper describes an innovative approach for mapping two tree species using a combination of commercial very high resolution WorldView-2 (WV2) images and Landsat time series data. As a first step, this study used a supervised object-based classification of WV2 images covering relatively small test sites distributed across the region of interest. Using these classification maps as training data, wall-to-wall mapping of fractional coverages of spruce and pine was achieved using multi-temporal Landsat data and Random Forests (RF) regression. The method was applied for the entire state of Bavaria (Germany), which comprises a total forested area of approximately 26,000 km 2. As applied here, this two-step approach yields consistent and accurate maps of fractional tree cover estimates with a spatial resolution of 1 ha. Independent validation of the fractional cover estimates using 3780 reference samples collected through visual interpretation of orthophotos produced root-mean-square errors (RMSE) of 11% (for spruce) and 14% (for pine) with almost no bias, and R 2 values of 0.74 and 0.79 for spruce and pine, respectively. The majority of the validation samples (75% (spruce) and 84% (pine)) were modeled within the assumed uncertainty of ± 15% of the reference sample. Accuracies were significantly better compared to those achieved using a single-step classification of Landsat time series data at the pixel level (30 m), because the two-step approach better captures regional variation in the spectral signatures of target classes. Moreover, the increased number of available reference cells mitigates the impact of occasional errors in the reference data set. This two-step approach has great potential for cost-effective operational mapping of dominant forest types over large areas. 1. Introduction Climate change is expected to have an important influence on future growing conditions that will likely have a great impact on tree species. Depending on the location, some trees species will benefit, for example, from increasing temperatures and thus, extended vegetation periods. However, at other locations, increased temperatures are likely to lead to water stress, and hence, greater vulnerability of some tree species to abiotic and biotic disturbances (Lindner et al., 2010). In Central Europe, several coniferous tree species are of particular concern: the vulnerability of tree species such as European spruce and Scots pine is particularly hig...

Optimal Input Features for Tree Species Classification in Central Europe Based on Multi-Temporal Sentinel-2 Data

“…Another benefit is the computation of importance measures which can be used for the evaluation of the input data and subsequent feature reduction. In this study, a recursive feature selection process using the 'Mean decrease in Accuracy' (MDA) was applied similarly to other studies [18,43,44]. More information about the algorithm and its advantages, such as the importance measure for the input variables and the integrated bootstrapping, can be found in the literature [16,41,45,46].…”

“…Studies covering larger regions by combining data with different spatial resolution have thus far only focused on a small number of tree species [18], respectively, on tree species groups [19]. Likewise, existing continental-scale forest maps such as the Copernicus high resolution forest layer [20], only distinguish broadleaf and deciduous forests.…”

Pattern recognition of forest canopy using the airborne hyperspectral data and multi-bands high spatial resolution satellite sensor worldview-2 data. A results comparison and accuracy estimation

“…Помимо преобладающего для каждого участка типа лесной растительности, который как правило соответствует некоторому кварталу и выделу карты лесотаксации, на участке зачастую могут содержаться и иные неосновные типы растительности, поэтому качество "эталонных" участков с точки зрения их включения в базу данных разработанной системы распознавания типов лесной растительности может заметно различаться. Так в развитие содержательной части работы (Козодеров, Егоров, 2013) (а речь там шла не только об автоматизации обработки данных) предложен метод построения диаграмм средних величин "интегральная яркость" -"параметр смещения" по всем спектрам с участка, который позволяет заметно улучшить отбор участков в качестве эталонов того или иного типа растительности и повышает качество распознавания типов растительной поверхности по гиперспектральным самолетным Представляет интерес как использование альтернативных методов распознавания (например часто используются статистические подходы), так и перенос разработанных технологий распознавания типов лесной и прочих поверхностей на применение спутниковых данных высокого пространственного разрешения, таких, например, как WorldView2 (Pu, Landry, 2012;Immitzer et al, 2012;Waser et al, 2014;Ramoelo et al, 2015;Karlson et al, 2016;Immitzer et al, 2016;Immitzer et al, 2018). Некоторые результаты многочисленных представленных в литературе исследований с использованием другой спутниковой аппаратуры дистанционного зондирования представлены, например, в работах (Guerschman et al, 2009;Eitel et al, 2011;Gärtner et al, 2016).…”

Real-Time Prediction of Crop Yields From MODIS Relative Vegetation Health: A Continent-Wide Analysis of Africa

“…Aggregate crop yields will also be more accurate than municipality-level statistics, since they have less variation geographically and in time. In other studies, coarse resolution imagery has been unmixed to detect sub-pixel changes [29,30]. Researchers have used this process to detect specific crops over a region or to analyze small-holder farms smaller than a pixel.…”