Forest disturbances are central drivers of ecosystem dynamics, impacting stand structure and composition, carbon flux, biodiversity, and economic value of forests [1
]. Ecological and economic impacts may vary strongly between different types of forest disturbance [4
]. Thus, information on the occurrence and distribution of disturbance agents is important to assess the carbon balance of forests [5
], impacts on biodiversity [8
], and economic consequences [7
]. Despite the ecological and economic importance of forest disturbances, there is a lack of spatially explicit information on forest disturbances at the agent level, especially covering larger geographic extents [12
]. This lack of knowledge hampers the development of process-based ecological models, which might help with predicting changes in disturbance dynamics under future climate change, as well as the development of adequate management strategies [12
Recent advances in satellite-based monitoring of forest ecosystem dynamics have demonstrated the potential for remote sensing applications to map forest disturbances at the agent level. In this regard, the Landsat satellite program constitutes a unique data source, allowing mapping forest dynamics at a stand scale over large geographic extents and long historic periods (>30 years) with relatively high temporal resolutions (annual or higher) [14
]. Following the opening of the USGS Landsat Global Archive in 2008 [15
], new methods have been proposed for exploiting the full potential of the high repetition rate offered by Landsat. That way, several automated change detection algorithms have been developed, which operate by fitting a simplified model to a time series of Landsat observations [16
]. Approaches based on Landsat time series have been shown to be very effective for detecting and characterizing different processes of forest ecosystem dynamics [21
While the detection of many forest disturbance processes using Landsat data may be considered operational given the diversity of successful applications in forest ecosystems around the world [14
], the discrimination of different disturbance causes (i.e., the attribution to disturbance agents) remains a prevailing challenge for remote sensing research [25
]. First studies have used the outputs of time series based change detection algorithms to attribute disturbance agents through classification models [26
]. Such approaches allow for the in-depth study of agent-level forest disturbance dynamics at regional to continental scales [28
]. However, current studies mostly mapped forest disturbances in North-America, where clear-cut harvest, fire, and large-scale insect outbreaks dominate as disturbance agents [26
]. Approaches developed in those forest ecosystems might be difficult to transfer to regions of lower disturbance intensity and smaller disturbance patch sizes, such as the forest ecosystems of Central Europe [32
In Central Europe, windthrow and bark beetles are the major natural forest disturbance agents [2
], which have been shown to interact in complex ways [34
]. Both levels of windthrow and bark beetle disturbances have increased in European forests over the last decades [2
], and are expected to further increase in response to climate change [37
]. Given the wide range of important ecological and economic impacts associated with both disturbance agents [5
], the monitoring of windthrow and bark beetle disturbances is critical to assist the sustainable stewardship of forest ecosystems in Europe [12
]. Thus, remote sensing-based monitoring of disturbance agents in Europe requires methods that can robustly identify and discriminate these two most influential natural disturbance agents. While there is a range of studies on Landsat-based characterization and identification of insect-related forest disturbances [27
], windthrow disturbances have received considerably less research attention [41
]. Further, most of Central European forests are under active management, where harvest—in particular selective logging (i.e., thinning or shelterwood cutting)–is an important disturbance agent [42
]. Also, stands disturbed by natural disturbance are commonly salvaged and sanitation felling is routinely applied to dampen the spread of bark beetles [43
]. These complex interactions between natural and human disturbances make the attribution of disturbance agents challenging.
Spectral-temporal metrics derived from Landsat time series have been successfully used to identify and discriminate different disturbance agents [27
]. Such metrics can be obtained from automatic change detection algorithms and provide key characteristics of spectral-temporal trajectories such as the spectral change magnitude or the duration of spectral change. While spectral-temporal metrics can deliver important insights into the characteristics of forest disturbances [49
], recent studies have demonstrated the potential of improving the identification of disturbance agents by incorporating additional sources of information [26
]. These additional information streams include context-based metrics, which relate the location of detected disturbances to landscape characteristics influencing the predisposition of forest stands towards different disturbance types (e.g., topography or forest type [26
]), as well as spatial metrics describing the spatial configuration of disturbance patches [26
In this study, we investigate the potential of detailed temporal information derived from intra-annual Landsat time series to improve the mapping of disturbance agents. Previous studies on disturbance agent attribution have used time series of annual best observation images [26
]. At the same time, great advances have been made regarding change detection methods including intra-annual Landsat data [20
]. For instance, Zhu et al. [20
] have developed a change detection algorithm working on all available Landsat observations that detects forest disturbances by comparing new observations against values predicted by a model fitted to stable forest pixels. Hilker et al. [51
] used a fusion model incorporating Landsat and MODIS data for detecting forest disturbances with high temporal and spatial resolution. As different types of forest disturbances often exhibit a distinct seasonality, such as windthrows [53
], insect outbreaks [55
] or fires [57
], information on the timing of disturbance events within a year may potentially help to infer the underlying disturbance agent [13
Moreover, combining Landsat time series data with additional information on natural disturbance events and/or their underlying causes may improve the attribution of forest disturbances. Previous studies combined climate data [59
] or independent disturbance observations from aerial surveys [60
] to predict insect outbreak dynamics. Here, we test whether including external information on the occurrence of storm events will help improve the identification of storm-related disturbances.
The aim of this study was to test Landsat time series-based predictors, context-based topography-related metrics, and prior information on wind disturbances for mapping forest disturbance agents in Central Europe. We tested our approach at three study sites featuring varying disturbances patterns and forest management intensities. Specifically, we wanted to address the following research questions:
How well can the major disturbance agents of Central European forests (i.e., windthrow, bark beetles, and harvest) be discriminated using Landsat-based agent attribution models?
Does incorporating information on the intra-annual timing of disturbances improve the attribution to change agents?
Does including prior knowledge about the occurrence of major storm events improve the attribution to disturbance agents?
Are the disturbance attribution models generalizable across study sites, i.e., are there differences in model performance and relative importance of predictor variables?