1. Introduction
Amplified climate warming in the Arctic [
1] is causing rapid ecological changes [
2], many of which may act as important feedbacks to climate change [
3]. In terrestrial ecosystems, sustained long-term increases in satellite measurements of the Normalized Difference Vegetation Index (NDVI) suggest widespread increases in vegetation productivity [
4,
5]. The positive NDVI trends, typically referred to as ‘Greening’, have largely been attributed to increases in the stature and extent of woody shrubs in Arctic tundra [
6,
7,
8]. However, nuanced heterogeneity in the ‘greening’ trends is occurring at spatial scales that are too fine to be resolved using medium- to coarse-resolution (e.g., 0.30–1 km) data with records long enough for trend detection. In fact, several recent studies utilizing data from Uncrewed Aerial Vehicles (UAV) have revealed that vegetation heterogeneity occurring at spatial scales of one meter or less is difficult to capture even with contemporary high- (~1–3 m) to moderate-resolution (e.g., 0.3–1 km) satellite data [
9]. Failure to capture this fine-scale heterogeneity means that satellite derived estimates of vegetation biomass and productivity may be underestimated by upwards of 10% [
10]. While this emerging research provides important insights about interactions between vegetation heterogeneity and observational scale in tundra ecosystems, there are large portions of the Arctic dominated by boreal forests [
11], where similar scale-related issues may exist [
12], but less research has been conducted.
Cajander larch (
Larix cajanderi Mayr.) forests dominate the landscapes across northeastern Siberia, and like tundra ecosystems, exhibit greening trends in long-term satellite data [
4,
11,
13]. In addition to enhanced productivity with warming, Cajander larch forests are likely experiencing increased wildfire activity that has the potential to alter post-fire vegetation recovery, carbon cycling, and ecosystem resilience [
14,
15,
16,
17,
18,
19]. Mature Cajander larch forests typically have low tree density, with understory vegetation communities similar to those found in tundra ecosystems [
20,
21] leading to a heterogeneous mixture of vegetation that makes it challenging to infer the ecological drivers of change [
22] or to link field data of vegetation composition and cover with satellite imagery [
12,
23,
24]. Moreover, recent research has shown complex interactions between fire severity and post-fire tree and shrub recovery [
25], and highlighted the possibility that recruitment failure can result in transitions from forests to grasslands [
26] that have important climate feedback implications [
27]. Consequently, the finer spatial resolution of UAV data (~1–10 cm) has the potential to discern more ecological heterogeneity for mapping and monitoring vegetation dynamics [
28,
29,
30,
31,
32,
33], which could be particularly useful for characterizing post-fire recovery [
34] as it relates to recruitment failure or shifts in forest structure.
The investigation of UAV data acquired from visible and multispectral imagery on post-fire landscapes is a burgeoning area of research. This may be particularly valuable in northeastern Siberia where aircraft campaigns are not feasible, short growing seasons with high cloud cover limit the availability of commercial high-resolution satellite imagery, and the remote nature of field sites means that field access maybe extremely limited. For example, the sites we describe in this study were located hundreds of kilometers from the nearest settlement and had to be accessed by boat. This meant that field time was extremely limited and that ultra-high-resolution UAV maps had the potential to provide additional invaluable information about these remote ecosystems. Recent research has focused on both the use of UAV data collected from traditional digital cameras solely in the visible spectrum using a single sensors (i.e., red, green, and blue wavelengths [RGB]) [
33,
35] as well as from more sophisticated multispectral sensors that includes data from multiple sensors in the visible and near infrared (NIR) wavelengths [
31,
32]. The Green Chromatic Coordinate (GCC), a greenness ratio reliant on the visible portion (RGB) of the electromagnetic spectrum, is a well-established metric commonly used in phenology studies [
36,
37] and has shown promise as a measure of vegetation recovery post-fire in other ecosystems [
35]. UAVs with RGB cameras offer a more readily accessible and cheaper solution than UAVs with multispectral cameras. Although less is known about how metrics like GCC that are calculated from RGB cameras correspond to vegetation in northeastern Siberia. Multispectral data allows for the use of vegetation indices such as the normalized differenced vegetation index (NDVI), which has been used extensively for monitoring and classifying vegetation with satellite imagery [
38,
39], but also has shown inconsistencies across satellite platforms [
40,
41]. For high-latitude ecosystems, cloud and snow-free high-resolution satellite imagery is particularly limited, making landscape-scale assessments challenging to execute [
42], and so UAV imagery may be an ideal data source for northeastern Siberia.
Here, we evaluate how well two common vegetation indices, GCC and NDVI derived from RGB and multispectral sensors, respectively, are able to predict field-based measures of post-fire vegetation (e.g., aboveground carbon biomass, live tree basal area, tall shrub basal area) across fire perimeters in northeastern Siberia. We examine how these relationships vary with spatial scale and when moving from the unburned edge into the burn perimeter in Cajander larch forests. We had three main objectives. First, we determined how GCC and NDVI relate to each other across a fine (25 cm) to coarse (10 m) spatial resolution to understand how fine-scale heterogeneity influences reflectance. Second, we evaluated how well these vegetation indices are able to predict aboveground carbon biomass and basal area of live trees and tall deciduous shrubs. Lastly, we assessed the differences in vegetation indices between burned and unburned sites and how vegetation indices change with distance from the unburned edge into the fire perimeter.
4. Discussion
We found that both GCC and NDVI derived from UAV data serve as useful indicators of post-fire vegetation conditions in Cajander larch forests. Our results demonstrate the utility of UAV imagery for quantifying fine-scale variation in vegetation dynamics in landscapes where field access and availability of high-resolution satellite imagery are limited. GCC and NDVI were most strongly correlated at coarser spatial resolution (e.g., 3-m, 5-m, and 10-m radii) compared to finer spatial resolutions (e.g., 1-m or less), which is likely due to minimizing the effects of background reflectance. Aboveground carbon biomass was predicted best by NDVI at 3-m, 5-m, and 10-m radii, which makes sense given the scale of field sampling and the fact that fine resolutions capture the heterogeneity of an individual plant rather than patches dominated by different plant functional groups [
10]. NDVI may perform better than GCC in predicting biomass because it is derived from radiometrically calibrated imagery, and includes near-infrared and red wavelengths that capture the spectral characteristics of plants. In contrast, the visible bands used to calculate GCC may be more sensitive to radiation reflected from non-photosynthetic surfaces. These findings highlight the importance of spatial resolution for linking UAV and field-based measures. NDVI was a stronger predictor of live tree basal area compared to GCC, and neither vegetation index was a strong predictor of shrub basal area. Both indices seem to distinguish between burned, unburned, and visualizations indicate a distinguishable edge effect. For GCC, the range of values across both burned and unburned matches observations in other vegetation studies [
57,
58]. For NDVI, the range of values across both burned and unburned is characteristic of landscapes dominated by shrubs or a mixture of shrubs and trees with an open canopy structure [
12,
59]. These findings illustrate the utility of UAV data for NDVI in this region as a tool for quantifying and monitoring the post-fire ecological response.
4.1. Correspondence between GCC and NDVI
The relationship between GCC and NDVI demonstrates the utility of GCC for characterizing spatial heterogeneity of vegetation in high-latitude Cajander larch forests. To date, there has been limited evaluation of spatial correspondence between GCC and NDVI. Vegetation indices derived from RGB and multispectral imagery are comparable for measuring phenology of plant green-up [
60]. Commercial, off-the-shelf UAV units with integrated RGB cameras are cheaper and easier to use than stand-alone multispectral cameras that must often be integrated with a UAV by the user. However, radiometrically calibrating RGB imagery is less straightforward, meaning that images may vary with illumination condition, and this may have contributed to the variability in our results. While UAV with built-in multispectral cameras are becoming more readily accessible, they are considerably more expensive than consumer UAV, such as the DJI Phantom 4 that we used. The GCC index provides a robust measure of greenness for measuring spatial and temporal variation in temperate conifer forests [
36,
37], Mediterranean pine forests [
35], and assessing forest stand vigor in North American sub-boreal systems [
61]. Testing the viability of vegetation indices calculated from RGB is timely since UAVs are becoming more financially accessible and can accelerate data acquisition in post-fire landscapes [
34,
62]. Our results demonstrate that the ability to acquire high resolution imagery in remote landscapes provides useful new opportunities to improve overall understanding of vegetation dynamics and provide valuable context for freely available coarse resolution satellite imagery.
A larger radius for pixel aggregation resulted in a better correspondence between NDVI and GCC, indicating that a coarser spatial resolution improves the signal to noise ratio. The finest spatial resolution (25-cm radius) may merely be capturing the reflectance from one component of the broader landscape like a patch of lichen/moss or a single shrub, while the coarser resolution (10-m radius) is aggregating the reflectance of these distinct vegetation components into a single vegetation index value. These landscapes have substantial variation in biomass, plant communities and structure that contribute to spectral heterogeneity. Aggregating pixels is considered an appropriate measure for minimizing the influence of vegetation heterogeneity on spectral index values [
10,
63] and relevant for evaluating linkages between field-based measures and UAV data [
10]. Aggregating pixels over a spatial area into a mean pixels value improves linkages to coarser resolution remote sensing imagery [
51], and spatial aggregations have been evaluated in several investigations e.g., [
10,
52,
53,
54]. This finding indicates that the spatial resolution is important, and the patchiness of these ecosystems influences each vegetation index. Continued research is needed to determine why GCC and NDVI agree less well at smaller spatial scales, with field observations at similarly small scales to determine the ecological drivers of this mismatch [
10].
4.2. Linking Field-Based Measures and Vegetation Indices
NDVI was a stronger predictor of aboveground carbon biomass compared to GCC. This finding points to the value of multispectral imagery from UAVs in this region. The 3-m, 5-m, and 10-m resolution demonstrated stronger relationships with field-based measures, pointing to the importance of scale [
63] and multispectral imagery for these ecosystems. The composition and structure of larch forests can be variable and patchy across the landscape [
20,
21] and contributes to spectral heterogeneity [
10,
12,
64]. This finding suggests that NDVI could serve as a proxy for aboveground carbon biomass, and indicates that commercial high-resolution satellite data (e.g., PlanetScope or Worldview) have sufficiently high spatial resolution to capture variation in biomass as well. However, previous evaluations on this landscape found no relationship between larch aboveground carbon biomass and NDVI sampled at high to moderate spatial resolution (3–30 m) using PlanetScope and Landsat imagery [
12]. This, in conjunction with our results, highlights the potential importance of understory shrub biomass, and suggests that UAV-imagery may serve as an important linkage between field-based and satellite-based measures of ecosystem characteristics related to vegetation [
10]. Targeted work aimed at understanding how relationships between vegetation indices and field measurements of vegetation vary with spatial scale and plant functional type is necessary to integrate linkages between UAV and satellite imagery.
NDVI predicted tree basal area better than GCC, while neither index predicted shrub basal area well. Forest landscapes consist of vertical and horizontal heterogeneity, and evaluating a single component of the vegetation structure may not necessarily be well represented by vegetation indices that are measuring all vegetation in an area. While tree or shrub basal area generally increased with increasing vegetation index values, this is expected given that photosynthetic biomass scales allometrically with the basal area for both trees and shrubs in these ecosystems [
20,
43,
65]. However, tree and shrub abundance only explained a portion of the variation in vegetation indices, and graminoids and non-vascular plants also exert a strong influence on NDVI at fine scales in ecosystems in this region [
59,
66]. This suggests that more detailed vegetation community sampling across fire boundaries, coupled with UAV surveys, may help to elucidate post-fire understory vegetation dynamics in these ecosystems.
Individual pixels from either UAV or satellite imagery represent a mixture of spectral signatures due to the heterogeneous structure of shrubs and trees, that from a top-down view are composed of leaves and woody material associated with twigs, branches, and stems. Further heterogeneity is introduced by spectral differences between patches of non-vascular vegetation or bare ground visible through gaps in tree and shrub canopies. This mixture of materials means that, at all but the finest spatial resolution, a single pixel captures characteristics that are a combination of surface types [
67], and represents a continued challenge for linking field data with remote sensing imagery [
12]. While finer spatial resolution has the potential to mitigate this issue, it will require field sampling at finer scales as well. Still, the heterogeneity will be a consistent component of optical remote sensing in these ecosystems as vegetation patches can be as small as ~50 cm in diameter [
59].
4.3. Postfire Vegetation Index Patterns Across Burn Perimeters
Both vegetation indices clearly differentiated burned and unburned areas in even the oldest fire scars, supporting the notion that larch forests are slow to recover and return to pre-fire greenness levels measured by reflectance [
20,
64]. These findings align with the general understanding that larch seedlings are slow-growing and often obscured by tall shrubs for one to two decades post-fire. The short growing season and slow growth of Cajander larch results in a protracted contribution of larch to total ecosystem biomass [
4,
64]. In comparison, tree biomass in North American boreal forests accumulates more rapidly [
68,
69], resulting in mappable forest recovery trajectories [
70]. Satellite-based studies have shown the recovery in greenness measured by vegetation indices takes more than a decade across high latitude regions of the northern hemisphere [
4,
64,
71]. However, this delayed recovery is measured with MODIS, a coarse spatial resolution (0.5–1 km) satellite, that minimizes variability in landscape heterogeneity captured by the finer spatial scale UAV data. In southern Siberian forests, post-fire recovery is linked with fire frequency, burn severity, and surface temperature anomalies, where more severe burns show greater spectral recovery rates [
72].
Declines in GCC and NDVI with increasing distance from the burn perimeter illustrate edge effects that persist for several decades post-fire that are detectable with remote sensing imagery. However, post-fire reflectance signatures are most likely dominated by understory vegetation because seedling regeneration is slow, and seedlings are quite small. The edges of fire perimeters serve as seed sources for areas within the burn perimeter [
73] and have been shown to influence plant assemblages in the North American boreal forests of Ontario [
74,
75]. Due to the monodominance of Cajander larch, these forests depend on trees that survive fire to serve as a seed source. It is feasible that greater seed rain would occur closer to the unburned forest edge in the absence of surviving trees within the burn perimeter. This proximity to seed source would create edge effects for larch that are potentially more prominent than those observed for other boreal forest trees where regeneration depends on serotinous seed sources or resprouting. However, because tree density in these forests is low, and larch recruitment typically occurs over several decades, it is likely that the edge effects we observed are related to understory vegetation recovery. An additional reason that NDVI and GCC might decrease with distance from the burned edge is that some of the burned edges were not discrete and there was a transitional zone with some live trees that survived the burn up to 25-50 m away from the primary burn edge.
Larch understory vegetation communities are similar to those found in tundra ecosystems, where recovery of vegetation productivity takes longer in areas with severe burning [
76]. Consequently, it is feasible that the edge effects we observed were related to increasing soil organic layer combustion (i.e., burn severity) [
77,
78] with distance from the unburned edge whereby vegetation in less severely burned areas along the edge of the fire perimeter recovers faster. Declines in seed dispersal with distance from the unburned edge provide an alternative explanation, although many common understory non-vascular plants are known to resprout from surviving underground parts [
76]. Continued research on spatial variation in burn severity, post-fire propagation, and edge effects are needed to elucidate drivers of reflectance gradients observed at the edge of fire perimeters in northeastern Siberia.
5. Conclusions
Our study illustrates useful relationships between field-based vegetation measures and UAV-derived vegetation indices across post-fire Cajander larch forests. We found that heterogeneity in vegetation indices varied with spatial resolution, and lead to differences in the ability of vegetation indices to explain variability in aboveground carbon biomass. The spectral distinctions between burned, unburned, and gradients across fire perimeters reveal edge effects associated with gradients of vegetation recovery that may be a consequence of post-fire ecological response, burn severity, understory vegetation dynamics, larch dispersal dynamics, or some combination of all these factors. Although NDVI generally corresponded more strongly with field observations, GCC also correlated with vegetation characteristics, highlighting the potential utility of easily accessible RGB sensors. Overall, our research demonstrates the utility of UAV data for documenting post-fire vegetation dynamics in Cajander larch forests, while also highlighting the need for continued research focused on understanding how relationships between vegetation and spectral indices vary across fine spatial scales.