Classification of pine defoliation by the common pine sawfly (I, II)

6.1.1. Tree-level classification (I)

Random forest classification was ran first for the scheme of DEF1 (two classes, threshold = 20% of defoliation) with all the 26 laser metrics to derive the scaled importance of the metrics. Mean return intensity was observed to be the most powerful predictor. However, because this metrics was not calibrated it was excluded from the following analysis. It is anticipated that dominant Scots pines are often more severely defoliated by the common pine sawfly than suppressed trees (section 4.2.). To classify defoliation level instead of tree size, Hmax and Hmean were also eliminated. Based upon the preliminary runs, the three most important features were h10, Hstd, and p70. These three metrics were used in all further classifications. The highest overall accuracies for defoliation classification were obtained with the schemes DEF2 (86.5%, SD ±6.1%) and DEF3 (85.4%, SD ±4.6%). As assumed, DEF5 with a highest number of defoliation classes (four) gave the lowest overall accuracy (71.0%, SD ±10.1%). In general, classification of healthier trees was more successful.

Nevertheless, most of the trees were classified at least to an adjacent class. In case of the thinned LiDAR data, classifications were done only for the classification scheme DEF1.

Random forest classification performance did not seem to be sensitive to the pulse density.

6.1.2. Plot-level classification (II)

Approximately a half of the sampling plots (n = 106) were healthy and the other half defoliated. Inspection of the LiDAR Hmean and Hmax showed some differences between these groups. The preliminary predictors were chosen based on available research (e.g., Solberg et al. 2006; Kantola et al. 2010, 2011; Solberg 2010), correlation coefficients, and preliminary modeling results (i.e., based upon biological plausibility and statistical significance). On average, trees on the defoliated plots were taller than those of on healthy ones. In order to minimize the influence of stand structure, potential explanatory features having a high correlation with tree size were excluded, i.e., Hmean and Hmax. Percentages of the LiDAR returns from the upper canopy (p60, p70, p80, and p90) are relative features and cannot be directly associated with the tree size. The same metrics were also identified as eligible classifiers for pine defoliation. The mean values varied significantly in Student’s t-test between the two defoliation classes (p = 0.00); on healthy plots, a higher number of LiDAR returns was reflected from the upper canopy.

Based on the preliminary random forest runs, the most important classifiers were proportions of the upper-canopy LiDAR returns of p80 and p90. These LiDAR metrics were used in the further analysis. Defoliated plots were classified with an accuracy of 84.3%

(kappa = 0.68). The same metrics were used in examining influence of LiDAR pulse density in assessment of pot-level defoliation. The classification result was not very sensitive to the pulse density, ranging seemingly randomly from 77.1% to 89.3%.

6.2. Landscape-level hemlock mortality (III, IV) 6.2.1. Effect of topography on hemlock mortality (III)

Totaling of 9,881 dead trees were identified within the watershed (Figure 7A). The vast majority of the trees was found in the northern part of the area, as well as reasonable nearby the river. The digitized canopy surface areas for potential gaps of the 1977 dead trees ranged from three m² to 88 m² (mean of 36 m²), with a positively skewed frequency distribution.

When the gap size distribution was generalized to the whole dead tree population, a canopy surface area of 7.2 ha was estimated, corresponding 0.1% of the investigated areas within the watershed.

There is a drastic North-South directional difference (~1000 m) in elevation within the area. The slope angles varied from flat (0 degrees) to very steep slopes of 80 degrees, facing all the directions. Distributions of topographic features of the dead trees differed from those of the whole study area (Figure 8). Dead trees were the most abundant in the high elevations of 900 - 1000 m and 1000.1 - 1100 m, as well as on slopes to north and northwest. Not distinct pattern for slope could be detected, however, a mild trend of increasing density along with steepening slope. Noteworthy is that much of the terrain in the area is on mild to moderate slope. A vast majority of the tree mortality occurred close to the Linville River; with distances ranging from zero to 2,760 m (Figure 9). Twenty five percent of the cumulative distribution of dead trees were located less than 60 m from the riverbanks (50% and 75% of were met at 94 m and 577 m, respectively).

Spatial pattern of the dead trees was analyzed for 54 sub-areas (Figure 7B). The spatial pattern was clustered at all spatial scales from one to 250 m in 36 sub-areas and random at all the scales in three (Figure 7C). Mixed spatial pattern, varying between random and

clustered, was detected in 15 sub-areas. No dispersed (even) pattern was observed. Clustered pattern of dead trees was mainly observed on higher elevations and in the proximity of the stream. The number of dead trees within the one km² sub-areas ranged from zero to 1192 (Figure 7D).

Figure 7. A) Identified dead trees within the Lower Linville River watershed, B) Subdivisions for the Ripley’s K analysis, C) Spatial pattern within the subdivisions, and D) the related dead tree density.

Figure 8. Frequency distributions related to topographic features of elevation, aspect, and slope in relation to A) density of dead trees within the study area, and B) for the whole study area.

Figure 9. A) Distribution of distances of the detected dead trees in the Lover Linville River Watershed, and B) A cumulative distribution of the distances of the dead trees. Black solid and dashed lines represent threshold distances of 25% (60 m), 50% (194 m), and 75% (577 m) of the cumulative distribution.

6.2.2. Mapping hemlock mortality within forest landscapes (IV)

The forested area was extracted within the extent of the LiDAR point cloud. The obtained classification accuracy was 93.5% (κ = 0.84). The resulting forest canopy cover, totaling 30.2 km², represented 76.9% of the classified area. The forest canopy cover was classified into dead trees, conifers, and hardwood with an overall classification accuracy of 98.1% (κ = 0.96).

Coverage by broadleaved species was the greatest of the classes, over 55% of the total classified area. Conifers occupied 42.6%, and dead trees 2.1% of the area. Dead trees represented over 0.6 km² of the forest canopy cover and conifers covered 12.9 km². The resulting classification image linked with a DTM showed that coniferous species were mostly located in drainages and on northern and western slopes (Fig. 10). Dead trees were mostly aggregated either in the proximity of the river, along drainages, or at higher elevations.

Figure 10. Classification image combined with digital elevation model at the Linville River Gorge. Classified conifer patches show as green and dead trees as red.

6.3. Insect-induced defoliation in Fennoscandia (V)

Five years was chosen as the number of years for the reference conditions. The ROC curve for EVI2 data resulted in an optimal z-score threshold of –2.9. With this z-score threshold, 50% of the damaged observations were correctly detected with no misclassification of healthy observations. The method was also tested with levels of 15% and 20% of defoliation.

The results indicated that the detection accuracy was not highly affected by the used defoliation threshold. The ROC curves of the NDVI data showed a zigzag pattern. This inhibited selection of the number of years for reference conditions, as well as optimal z-score threshold. Because use of NDVI also resulted in extensive misclassification of healthy plots with low TPR rate, EVI2 was assumed to outcompete NDVI in this sub-study. However, the preference may depend on the goals of a monitoring task.

Evaluation was done for the analysis with EVI2 data only. In Outokumpu, 50% of the defoliated stands used in the evaluation were detected with a misclassification rate of 22%.

Only 27% of the damaged stands were detected (misclassification rate of 54%) in Ilomantsi.

To demonstrate the ability to tailor the method according the purpose, the z-score threshold was adjusted. A higher threshold (z = –2.1) was applied. This resulted in misclassification of the healthy stands of 50% and 35% for training and testing data set respectively, in Outokumpu. In Ilomantsi, the corresponding rates were 46% and 70%. In Abisko, six years was assigned as a number of years for reference conditions with the z-score threshold value of -6.0. The point closest to (0, 1) suggested a detection rate of 75% with a misclassification of healthy sampling units of 19%.

6.4. Projecting potential distribution of the hemlock woolly adelgid (VI)