Remote sensing allows for the broad-scale monitoring of forests around the globe using precise spatial information and frequently repeating observations to identify and detect changes in vegetation presence, abundance, and condition. These datasets lead to greater knowledge of forest ecosystem patterns over large spatial extents [1
]. Remote sensing has a long history in aiding forest inventory, including the identification of forest cover in the form of geospatial maps, assessment of forest health as related to defoliation and discoloration [2
], and quantitative models that provide predictions such as timber volume, basal area, aboveground biomass, and carbon [3
]. Derived maps and reports are used to enhance on-the-ground measurements, identify areas of interest for further investigation, and provide consistent, wall-to-wall data coverage to inform land management decisions [4
The ash genus (Fraxinus
spp.) is widely distributed across much of central and northern North America both in urban and traditional forest environments. Since the identification of the invasive emerald ash borer (Agrilus planipennis
Fairmaire; EAB) in 2002 in Detroit, Michigan, USA, there has been widespread mortality of ash across North America [6
]. Currently, EAB is present in 35 U.S. states and several Canadian provinces. In Minnesota, USA, EAB is locally present in several counties in the central and southern regions of the state. The EAB larvae feed on the phloem of ash trees and contribute to tree mortality within one to five years of infestation. After introduction of EAB, annual ash mortality can increase by as much as a 2.7% in a county [7
]. The anthropogenic movement of the beetle through infested firewood or other products has aided in the introduction of EAB to new, distant locations [6
Ash is a prominent tree species in Minnesota forests that exists across many land cover types. Miles et al. [8
] estimated that there are 1.1 billion ash trees in Minnesota that are at least 2.5 cm in diameter at breast height (DBH) or greater, accounting for 8% of all trees in the state. However, due to its low economic value, there has been limited research on this forest type, including basic forest inventory to understand the extent and locations of ash stands. The three species of ash found in Minnesota are black (Fraxinus nigra
Marshall), green (Fraxinus pennsylvanica
Marshall), and white ash (Fraxinus americana
L.). The majority of the individuals are black and green ash with relatively few white ash trees. In Minnesota, black ash is dominant in two vegetation types: (1) Black ash-elm/trillium vegetation communities that occupy moist sites with deep organic soils and (2) black ash/yellow marsh marigold vegetation communities on sites with better drainage. Black ash are dominant in the overstory of swamp forests that occur in extensive complexes, in topographically low depressional areas, and at transitions between upland forests and peatlands [9
]. However, black ash can be found in lower abundance in nearly all vegetation types where it mixes with other species [10
Historical disturbances have recently been mapped using the full extent of the Landsat archive across Minnesota’s forests, highlighting the value imagery time series to monitor forest dynamics [11
]. Remote sensing at the individual species level is difficult in mixed-species forests like Minnesota’s because of the numerous species and high degree of heterogeneity in small areas, i.e., the size of stands where forest management decisions are primarily made. Fortunately, ash has certain biological and physiographic features that may help to differentiate it from other tree species. Deciduous phenology patterns are useful for speciation at a broad level (e.g., deciduous vs evergreen). Ash species drop their leaves earlier than other deciduous species in the fall [12
]. This is useful from a remote sensing perspective because the change in optical reflectance from healthy, green leaves to barren, grey branches can be an indicator of species. However, differences in local site conditions and climate influence phenology as well, and additional information is likely required to accurately identify species. A physiographic characteristic of ash, particularly in black ash, is the prevalence of the species in relatively low elevation positions with high degrees of soil moisture such as swamps, bogs, fens, and other forested wetlands. These characteristics separate ash from other deciduous tree species like aspen (Populus
spp.), one of the most common hardwood trees in Minnesota.
Given the extent of ash and the impending EAB threat in Minnesota, there is a need for accurate and high-resolution maps of ash presence and abundance. These maps will help to quantify the current extent of ash for future analysis and develop land management plans for areas of high EAB infestation risk. Previous maps of ash abundance have been generated at coarse spatial resolutions, i.e., 250 m [10
], using alternative imputation methods [13
], or within single Landsat scenes [14
], but not at a moderate satellite spatial resolution for the regional extent. The objective of this work was to produce a 30-m resolution map of current ash presence/absence in Minnesota. Ash abundance was modeled in terms of basal area using lidar height metrics.
The detection of ash species using time series at 30-m spatial resolution using composite imagery and cloud-computing software like Google Earth Engine provided a new possibility to assess ash presence. Combined with lidar data, ash abundance was determined across Minnesota in units of basal area. Distinguishing individual species from optical imagery is notoriously difficult, largely because the spatial resolution of imagery causes specific vegetation signatures to become mixed with other species [30
]. This is especially true in heterogeneous environments such as the Laurentian mixed forest where ash is more abundant. Our approach using NDVI metrics to account for the unique phenology of ash helped to address this limitation and distinguish differences in vegetation. Imagery with higher spatial resolution would reduce this issue but there are currently no such time series with sufficient duration.
The utility of large datasets, such as the Landsat archive, depend on efficient computation ability. Pixel-based time series modeling using hundreds of NDVI images would not be possible using traditional geospatial processing methods. Google Earth Engine facilitated the dense time series analysis through the use of existing cloud masking and reflectance correction that would have otherwise been a barrier to model development. The spatial resolution of the imagery was not adequate to observe individuals, but the phenologic pattern derived from the time series seemed to have captured species-specific patterns. Whereas the time series from different sensors encompasses a wide range and differing timespans, the number of observations from each sensor time series seemed to have greater impact than the “currentness” of the time series. Observations of seasonal fluctuations in vegetation describes ecological patterns that added value to the predictive model in all cases.
Whereas time-series imagery is beneficial to examine forest disturbances that occur quickly at a high magnitude [11
], e.g., fire or timber harvesting, and damages from insects, a relatively slow-acting change agent is also possible to examine using Landsat data [31
]. The application of these data serve a broader understanding of where ash forests exist in Minnesota and how much impact EAB will have on the state’s forest resources. Our detection accuracy of black ash (72%) aligns well with the accuracy rate of 85.5% observed by the authors of [14
] for the species in a single Landsat scene in northern Minnesota and the rate of 89.5% observed by the authors of [32
] across the northeastern Minnesota region. Our slightly lower detection accuracy compared to these studies was likely the result of the broader geographic area examined (i.e., statewide) and the difference in dates between lidar acquisition (2008–2011) and the field inventory measurements (2014–2018). Similarly, historical Landsat data are older compared to Sentinel-2 data, which also differ in acquisition dates from field inventory measurements. Compared to previous estimates of the extent of ash in Minnesota, our value of 1.25 million hectares aligns well with the design-based estimate of ash abundance provided by the FIA program. The FIA program estimates that ash is at least 50% of the total live tree volume on 0.45 million hectares of forest land but is a component of 1.7 million additional hectares of forest land where it occurs with other tree species [33
]. The distribution and abundance of ash observed in this analysis also generally agree with the statewide results of Wilson et al. [13
] and Kurtz et al. [10
]. Our approach was most similar to that of Engelstad et al. [14
] through the use of lidar, Landsat, and soils information as input data, while other studies modeling black ash have relied on Landsat and radar [32
] and the MODIS Terra data product [13
]. Our findings support the use of GEE as a useful tool to integrate remote sensing datasets of different resolutions. As indicated in Table 3
and Table 4
, a longer time series was associated with higher relative accuracy compared to shorter time series. In particular, the use of polygon-based forest inventory data (i.e., the FIM data) provided excellent training data to detect ash presence across contrasting landscapes in Minnesota’s diverse forest communities.
The spatial resolution of the classification map provides unique insights to the connectivity of forest cover types that is useful for targeting areas for strategic forest management. Geospatial information is ubiquitous in land management and landscape planning. The benefit of continuous imagery across boundaries circumvents some of the limitations of field-based measurements but does not replace the need for ground truth observations. This work provides a baseline of current ash abundance to attempt to quantify the risk of EAB infestation. The use of time series that date back to 1974 offers not only a more rigorous dataset, but also a window into where ash forests previously existed. It is possible that classification errors occurred due to the fact that older ash stands observed in historic time series have been succeeded by new cover types or transitioned to different species or a new ecosystem state associated with ash dieback [34
]. These classification errors may be more prominent within the limited regions of where EAB has been observed in Minnesota since 2009 (i.e., in central and southern Minnesota).
The complexity of species composition in Minnesota forests make remote sensing techniques difficult, indicating that additional studies may be required to improve the accuracy of satellite time series analyses. More accurate models may be generated in these additional studies with increases in temporal and spatial resolution. The limitations of 30-m pixels are known to underrepresent highly mixed forests [35
], and this would include ash or forests with a high understory density of ash trees. However, moderate resolution (i.e., 30 m) is likely sufficient to identify critical hotspots or connectivity analysis. The RandomForest classification probability was a highly skewed positive distribution where the majority of probability estimates were less than or equal to 0.1. Both the ash presence and absence error were minimized when 0.1 probability of ash classification was used as the cutoff threshold for ash presence. However, it is important to note that the relative abundance of ash on a plot will impact the detection of ash. Lower abundance of ash is more difficult to detect and, in these instances, the prevalence of misidentification of ash would increase.
Further research is required to understand how ash forest connectivity influences EAB dispersal capabilities. In particular, lidar-derived metrics such as the Compound Topographic Index are essential in determining dispersal capability as reflected in the presence of susceptible host trees. A wall-to-wall map of forest cover types and species distribution can guide selection of these target areas for further investigation or management. In addition, and in particular with EAB, the time of insect arrival (if known) should also be considered because decreases in ash abundance generally begins six to seven years after EAB is first detected [7
The detection of landscape change using satellite image time series is an instrumental tool for understanding forests. The access to free data such as Landsat and Sentinel and open research platforms such as GEE facilitates exponential growth in the realms of both research and management. Invasive species, impacts from climate change including shifting disturbance regimes, and the incorporation of multiple goals and objectives in forest management requires data on the composition and structure of forests. The continued evolution and merging of remotely sensed data with on-the-ground forest inventory data can aid managers in developing landscape-level plans. Whereas forest management often happens at a stand scale, the disturbance event (in this example, EAB), will impact the landscape and will cross ownership boundaries. County-, state-, and landscape-level maps promote communication across forest ownership boundaries. These mapping efforts facilitate the development of management plans that increase the health and resilience of forests.