3.1. Tree Species Classification
The two-step classification results are displayed in Figure 5
. In the unsupervised forest/non-forest map (Figure 5
a), the non-forest classes are mostly herbaceous grasses clustered in valleys and low-elevation wetlands. During our field trip, we learned from local foresters that many low-elevation areas were covered with natural forests before the 1987 fire. Trees failed to grow back in these lowlands because of high moisture conditions in soil (Mr. Huadong Wu, Vice Director, Fire Prevention Office, Tuqiang Forest Bureau, personal communication, 5 June 2015). In Figure 5
b, there are more Xing’an larch distributions in the unburned south. The regrown tree species in the north was dominated with white birch. Mongolian pine was mostly clustered in the southeastern end of the study area, where the climate is transiting to a warmer continental temperate climate.
In the confusion matrix (Table 2
), the overall accuracy of classification reaches 86.16% with a Kappa value of 0.80. Non-forest covers were optimally classified from satellite time series because of their distinctively different reflectance from tree species. As indicated in the overlaid error bars among the three trajectories in Figure 3
, high confusion of tree species was observed in the classification results. As an evergreen coniferous species, Mongolian pine had different spectral trajectories from other tree species at the start and end of growth, which resulted in a high producer’s accuracy of 94.12%, although its accuracy (72.73%) was low, which may come from its small sample size. White birch and Mongolian pine were less confused. Xing’an larch, however, turned out to be confused with both tree species. One possible reason was the mixed re-growth of all tree species after the 1987 fire as recorded in the 2010 NFRI inventory data. Studies have observed that white birch outcompeted Xing’an larch in severely burned forests [4
]. The mixed composition of tree species also had its natural causes, as the study area was transiting from boreal to warmer temperate climates.
Only nine Landsat images were available to build the time series in the study area. If more frequent observations were available, the confusion among tree species could be reduced from better defined trajectories in a growing cycle. Reference [9
] extracted trajectories of the post-fire stand regrowth index from the 15-day AVHRR Global Inventory Modeling and Process (GIMMS) NDVI product in 1984–2006. Its 8-km grid size, however, is too coarse to classify individual tree species in this study. More recently, data fusion techniques have been developed to build Landsat-like, daily image series by taking advantage of image pairs between coarse-spatial, but high-temporal resolution imagery (e.g., MODIS) and medium-spatial resolution Landsat data [26
]. While inter-calibration of different sensors could be a concern in the fused NDVI time series, it could be leveraged by explicit trajectory smoothing. The more frequent, Landsat-like time series could provide better insights for trajectory-based tree species identification.
The CART decision tree approach for tree species classification heavily relied on training samples for optimal threshold selection in decision rules. Training samples in this study were collected from surveyed pure polygons of forest inventory data. Uncertainties of inventory records could rise from field surveyors’ experience and the physical accessibility of the remote, natural forest in our study area. To leverage this, we collected a large amount of training samples (e.g., 189 samples for larch and 187 for white birch), although Mongolian pine had a much smaller training set due to its limited number of pure unit polygons in the study area (as shown in Figure 2
). Large sample sets improved the validity of thresholds selected by the CART approach. As demonstrated from the error bars of trajectories in Figure 3
, however, high uncertainties are expected in classifying the three tree species.
Given the abovementioned uncertainties, Xing’an larch reached accuracies around 70% in producer’s and user’s assessments. However, Figure 5
b represents the first spatial layer of Xing’an larch in the Greater Xing’an Mountains after the 1987 fire. Its spatially-heterogeneous distributions in the study area could reflect fire disturbances to its recovery in the past 30 years.
3.2. The NBR Fire Intensity Map and Xing’an Larch Re-Composition
The 1987 fire swamped half of the study area in the north (Figure 6
). Most of the burned areas had the highest fire intensity (rank = 1, highlighted in red), reaching 13.97% of the study area (Table 3
). Areas with other burned ranks were scattered around. Unburned forests, for example in the south of the study area, were dominated with Rank 10 (in bright blue), covering 53.35% of the study area. To demonstrate the relative severity of fire damages, we grouped the 10 ranks of NBR into high, medium, low and unburned. Details about the ranks and their associated fire intensity categories are listed in Table 3
Valleys and low-elevation areas in the south were not burned, but were also characterized with lower NBR values due to their spectral differences from forests. They have Ranks 7–9 in Figure 6
and could be misclassified as low-intensity burned areas using the categories in Table 3
. For this reason, we used a cutline at the south edge of fire scars (marked in Figure 6
) to separate burned area in the north and unburned area in the south. In the south of this cutline, all ranks lower than 10 were assigned 10 to represent unburned areas. Areal percentages in Table 3
reflect the re-assigned ranks.
Tree species compositions in each NBR rank were summarized to statistically evaluate their regrowth under various fire intensities (Figure 7
). Only forested areas were calculated. A higher larch composition approximated its better recovery in burned areas. Figure 7
reveals a linear increasing rate of recovery of Xing’an larch under high to low burning intensities. In high-intensity burned areas (rank = 1), almost all trees died from the 1987 fire, and therefore, tree species were mostly grown from new seedlings [25
]. White birch outcompeted other species and dominated these areas with a high composition of 76.01% in forests. Xing’an larch could not compete with white birch in its early stage of post-fire succession; therefore, its composition dropped to 17.52% under high burning intensities. In areas closer to roads and Tuqiang Town, we observed patches of planted Xing’an larch during our field trip. However, with the lack of labor and budget, planted stands in the study area were very limited. Forests in natural recovery were thus dominated with white birch. Under medium burning intensity (rank = 2, 3, 4, 5), some trees were able to survive. Xing’an larch had a relatively higher composition and white birch had a lower composition than in high-intensity areas. Under low intensity (rank = 6, 7, 8, 9), trees survived the 1987 fire, and therefore, Xing’an larch composition was much higher than that in low-medium intensities.
For unburned forests (rank = 10) in the south of the study area, Xing’an larch, Mongolian pine and white birch covered 40.59%, 28.78% and 30.63%, respectively (Figure 7
). It fairly demonstrated the compositions of pre-fire natural forests in this region, i.e., higher Xing’an larch distributions in the south end of boreal Eurasia. It is also shown in Figure 5
b and Figure 7
that Mongolian pine has an un-proportionally high composition in unburned forests in the south, which may reflect tree species transition in a warmer climate below the boreal zone.
indicates that, within 30 years after the 1987 fire, Xing’an larch in the study area has not fully recovered. It composed 17.52%, 26.20% and 33.19% of burned forested in high, medium and low intensities, which had not reached its composition in unburned forests (40.59%) even with a higher distribution of Mongolian pine in the south. Different compositions of Xing’an larch in Figure 7
are supported by field studies in the past. In high-intensity burned areas, larch germination was heavily hindered by the lack of surviving seeds, while white birch rapidly took over via sprouting [3
]. Its higher composition rate in low-intensity areas could be attributed to the improved seeding conditions. After ground canopies were burned down, seeds from surviving larches had a higher chance to reach soil.
Aside from natural succession of forests, post-fire tree species composition was highly affected by human interferences. Shortly after the 1987 fire, salvage logging was widely operated by the Greater Xing’an Forest Bureau across the burned forests to maximize the profit and to prevent the consequent insect damages. Since Xing’an larch holds high economic value, a large volume of healthy trees in areas with lower burning intensities were also logged. These activities further reduced larch compositions in burned areas, but were not documented in open data records. The planting strategy in the region after 1987 shifted from timber harvesting to planting coniferous seedlings. However, the planting density varied, and only 10% of the burned forests were managed with planting [7
]. Since the early 2000s, the Natural Forest Protection Project named “Tian Biao” has been enforced, and logging activities were prohibited in the Greater Xing’an Mountains. Within a rigid fire monitoring network, there have been no severe fires in the study area since 1987. Some research actually concerned that intensive logging and successful fire suppression efforts have led to fire fuel buildup. Fires in this region become less frequent. Once broken out, however, they could be more intense than historical regimes.
In short, this study performed tree species classification from satellite time series relying on their unique trajectory patterns of spectral indices in a growing cycle. Past remote sensing applications in the study area often utilized limited scenes of satellite imagery for forest classification. While forest/non-forest types were fairly discriminated, detailed tree species composition cannot be effectively extracted because of their spectral similarity. With vegetation indices, various studies reported that greenness of boreal forests can grow back to pre-fire conditions in 5–20 years depending on fire severities and imagery resolutions [9
]. This study revealed, however, that Xing’an larch re-composition had not been fully recovered in a 30-year span. Boreal forests grow slower than the temperate biome. They may need a much longer time to reach a later stage of succession (e.g., mature forest stands) or have been changed permanently in this vulnerable boreal forest. This study only examined the most recent growing cycle using imagery in 2013–2015 to extract the current re-composition of Xing’an larch. As demonstrated in the AVHRR GIMMS-extracted stand regrowth index trajectory in [9
], image series of past cycles deserve further examination to understand the temporal path of larch re-composition in the long period after 1987.