1. Introduction
Boreal forest ecosystems normally are highly dynamic and resilient to a variety of changes which promotes stable development of forest stands across broad temporal and spatial scales. Normal variations in annual growth patterns and needle biomass need to be distinguished from disturbances leading to declining forest health. Disturbance interrupts successional development of forest ecosystems, affecting resources, the physical environment, population structure, and, in extreme cases, changing the direction of successional processes [
1,
2]. Forest disturbances can appear in different forms such as abiotic (storm, drought, frost, snow, fire) or biotic (pest insects, diseases, mammals) damages [
3,
4], causing threats to sustainable forest management and economic return in the Boreal Zone [
5,
6].
Climate change and increased extreme weather patterns can magnify the intensity of forest disturbances, altering the geographical range and productivity of forests, especially at higher latitudes [
7,
8,
9]. During global warming, increases in stress factors and patterns of insect outbreaks have been predicted [
8,
9,
10,
11,
12]. Due to rapid responses to elevating temperatures, pest insects can flexibly change their survival, development, reproduction, dispersal and geographic distribution [
13,
14,
15,
16]. Increasing numbers of pest insects have already begun to expand their normal geographic ranges, either pole-ward latitudinally or upward altitudinally [
17,
18,
19], or change their pest status within their ranges [
20]. Increased outbreak frequencies and spatial scales of forest pests evidently have already undergone changes during the most recent decade in Finland, particularly with pine sawflies [
20].
Development of modern, cost-efficient monitoring methods for forest sites affected by climate-driven disturbance agents is urgently needed [
21]. Monitoring of needle defoliation has typically been based on field sampling [
22], which consumes vast resources, only to yield results that may still be biased. Furthermore, estimates of future defoliation and yield losses are only qualitative. The Finnish Forest Research Institute carries out the National Forest Inventory (NFI) [
23], in which information on forest health is collected as a side product and monitored on a coarse level. The annual requirements of precise information on forest disturbances are not met by the current practice of forest health monitoring.
Remote sensing is an efficient tool for detecting changes in forested areas, such as disturbances [
24]. Current development in active remote sensing technologies, especially airborne laser scanning (ALS) techniques have resulted in new methods for carrying out various forest inventory tasks. With the capability of direct or derived measurement of forest structure, including canopy height, crown dimensions and above-ground biomass, ALS can be also applied for monitoring of forest hazards. Previous studies have shown that ALS data can be used to estimate several forest inventory attributes, such as the tree-, plot- and stand-level characteristics of tree height [
25,
26,
27], biomass [
27,
28,
29,
30], volume [
31,
32], basal area [
33,
34], tree species [
29,
35,
36] and forest operations [
37,
38]. ALS can be useful in projecting, detecting and monitoring forest hazards and tree defoliation due to its ability to directly measure vegetation structure [
39,
40,
41]. Recent studies and developments in methods have achieved more accurate ALS-based biomass detection [
39,
42,
43,
44,
45,
46,
47,
48]. Single trees biomass and defoliation level are highly correlated (e.g., [
49]).
The objective of this present study was to test the accuracy in tree-wise classification of needle defoliation after consumption by pine sawflies over a period of several years in a row. Defoliation estimations were made using several needle loss category classification schemes, in order to investigate their effects on accuracy. Classifications were based upon statistical metrics extracted from ALS data at the level of a single tree crown. The hypothesis was that the distribution of laser returns from defoliated trees differs from that of healthy, undefoliated trees. Kantola
et al. [
39] earlier investigated how to separate healthy and severely defoliated trees using an approach similar to that used here combining ALS data with high resolution aerial imagery. Vastaranta
et al. [
48] developed an area-based approach for mapping healthy and defoliated Scots pine stands with ALS data. In both of these studies, simple two-class defoliation classification schemes were used. This study included multiple higher level defoliation classification schemes in order to extract more detailed information. An additional objective involved determining the effect of laser pulse density on the classification accuracy.
4. Discussion
Forest inventory, mapping and monitoring methods have rapidly developed in recent decades. New methods are more often based on RS applications, especially using ALS. While the methods are changing and forest disturbances are becoming more abundant, there is an urgent need for new methods to map and monitor forest health. In the present study, statistical ALS metrics were tested in the classification of individual tree defoliation. The RF method with three selected ALS metrics was applied to estimate the accuracy of different combinations of defoliation categories and varying pulse densities.
A total of 701 trees were allocated in five different needle loss category classification combinations having 2–4 defoliation classes. The hypothesis was that a larger portion of ALS hits would penetrate deeper into tree crowns and the statistical ALS metrics differ between healthy and defoliated trees. The RF method showed some promising results using 2–3 defoliation classes, while use of further classes resulted in predictions that were only moderately accurate. However, RF was able to classify most of the trees in every classification scheme at least to the adjacent defoliation class. By analyzing classification accuracy of ALS data that was randomly thinned into 10 different pulse densities (2–20 pulses/m2), we found that RF classification was not overly sensitive to varying pulse density and that the overall classification accuracies did not vary considerably between different pulse densities.
In several studies, ALS data has been used in the estimation of forest characteristics other than defoliation at both stand- and tree-level (e.g., [
30,
57,
66,
67]). The utilization of ALS has been less studied in the field of forest disturbances. For example, Vehmas
et al. [
68] detected deadwood through canopy gaps, using ALS data [
37]. Solberg
et al. [
42] compared leaf area index (LAI) with ALS data in defoliated Scots pine stands. Kantola
et al. [
39] tested ALS data combined with aerial imagery for defoliation estimation by RF with two defoliation classes, and obtained an overall classification accuracy of 88.1% (kappa value 0.76). For comparison, Kantola
et al. [
39] also tested defoliation detection accuracy using only the ALS metrics (classification accuracy 80.7%). Their results with spectral features were slightly better than in the present study, but they used only two distinct defoliation categories,
i.e., healthy and heavily defoliated trees. In the present study, all the trees in addition to those having a threshold value were used in the analysis. In addition, the results were also fairly accurate with more than two classes. Vastaranta
et al. [
48] studied plot-level needle loss prediction for the same study area. They obtained an overall classification accuracy of 84.3% for two classes.
The defoliation level in the field was visually estimated, using the same procedure as the NFI of Finland [
23]. However, visual interpretation could easily have caused deviation in the results if the surveyors were not professionals. Naked-eye calibration is essential when two or more researchers are estimating the critical variable. Observers should also be able to distinguish a between years and within year natural variation in foliage biomass. In addition, prevailing conditions could have also caused bias in the defoliation assessment, such as weather, brightness, heavy wind, high tree density, and difficult terrain. Visual needle loss assessment was done with 10% accuracy and there are uncertainties in assessment. Due to these uncertainties, using narrower class limits is not justified.
Most of the trees were classified from ALS data into the correct or adjacent defoliation class. Misclassifications may have originated from the sensitivity of the reference visual defoliation estimation to errors. For example, a tree having an approximately a defoliation level of 20% could have been visually classified into classes of 10% or 30%. Development of better methods for needle loss estimation in the field may also improve the detection rate from ALS data.
All data were collected from trees in the same study area. A typical feature of the study area was that the taller and older trees in the dominant canopy strata were more heavily defoliated than the shorter and younger trees. This pattern is typical of
D. pini outbreak dynamics. Ovipositing females prefer the uppermost parts of the crowns, due to higher carbohydrate synthesis in these needles than found under more shaded conditions [
20,
69]. The high carbohydrate content, particularly of soluble sugars, promotes the survival of the following sawfly generation.
D. pini attacks suppressed understory pines only after completely consuming needles of taller trees. To avoid classifying tree height instead of defoliation, pure height features, such as
Hmean and
Hmax were not used in classification.
The result of the first RF run with all 26 ALS metrics indicated that the mean return intensity could be a powerful predictor of defoliation. In theory, intensity based upon wavelength of 1064 nm in near-infrared area, should differ between healthy and defoliated trees. In practice, the use of intensity is often problematic, because it has to be calibrated. In this study, the intensity was not calibrated. However, the power of raw intensity was also tested. When penetration (pene) was used together with intensity (int) inl classification, overall accuracies of 81.74% and 83.59% were obtained for two classes (DEF1 and DEF2). Based on this result, it could be assumed that a full waveform ALS could allow more accurate classification of defoliation.
The distribution of defoliation levels among trees was uneven in the study area. The number of trees suffering from heavy defoliation was quite limited, due to low daily temperatures in summer 2008. The classification accuracies were higher in healthier tree classes, which may have resulted from the scarcity of heavily defoliated trees. A larger proportion of trees having severe needle loss could improve the classification accuracy.
Recent studies have shown that the distributions of ALS features vary among different site types [
70]. In the present study, the site types were not taken into account because they varied only slightly within the study area. The distribution of ALS metrics probably varies depending on the size and hierarchy level of the trees. For example, Korpela
et al. [
30] found that dominant Scots pines had approximately 5% higher mean return intensity than the intermediate trees. In the present study, the smaller trees from the suppressed canopy cover level were excluded, but there is still considerable variability in the data.
Results of this study suggest that it may be possible to detect trees with differing levels of needle loss, although the detection accuracy showed less success with increasing numbers of defoliation classes. The results of the present study can be useful for improving the use of ALS in detection and mapping of damage by defoliating insects, which is usually a rare and clustered phenomenon. This study may also support further development of methodologies for inventorying defoliating pests. For example, with remote sensing data, the stratification could be carried out by focusing on areas where pest damage could be detected from preliminary remote sensing data. Suitable class intervals could be set at reasonable threshold levels to obtain adequate estimates, depending upon the nature of the forest disturbance in question.
To the best of our knowledge, the use of ALS-based ITD inventory for estimating tree defoliation has not been widely investigated. However, the results of this study are in some ways comparable with other studies using RS data in needle loss estimation. Ilvesniemi [
71] used the same Palokangas study area that was utilized here when investigating the use of aerial photographs and Landsat Thematic Mapper (TM) data in classifying defoliation of Scots pine at the plot level. The classification accuracies for features extracted from aerial photographs varied between 38% (nine classes) and 87.3% (two classes). The best explanatory variable for needle loss was maximum radiation of the near-infrared (NIR) channel in aerial images (
r2 = 0.69). Classification results with Landsat image features were slightly poorer (accuracies between 25.4% and 88.7%). Aerial images have also been utilized in other studies to detect tree-wise defoliation such as by Haara and Nevalainen [
72]. Their results showed that the tree-wise classification accuracy for reference data of Norway spruce (
Picea abies (L.) Karsten) was 68.9% with four classes.
Karjalainen
et al. [
73] used multitemporal European Remote-Sensing Satellite 2 (ERS-2) and Environmental Satellite (Envisat) satellite images and calculated the synthetic aperture radar (SAR) backscattering intensities (squared amplitude) of 400-m × 400-m grid cells. These SAR features were used to estimate defoliation (same two classes as used here). An overall classification accuracy of 67.8% was obtained, when 30% of the field reference was used in training and 70% for testing the model.
Vastaranta
et al. [
48] also studied the effect of pulse density for mapping plot-wise defoliation and their results were similar to this study. No remarkable sensitivity for pulse density was found in prediction. According to Kaartinen
et al. [
66] the pulse density may not affect the individual tree detection. In this study, the tree identification was only done with full pulse density data (~20 pulses/m
2).
The present study is one of the first steps towards developing an ALS-based system for monitoring changes in forest health (defoliation) in Finland. Optimally, defoliation mapping should be adopted in current annual practices. For example, it should be part of NFIs or operational forest management planning based on ALS inventory. Field surveys could provide information for growing stock estimation, precise information on defoliating pest agents and also coarse data on needle defoliation. Then, ALS data can be applied on demand to create maps for stem volume and defoliation status where precise information is needed.
Further studies are planned to focus on more heterogeneous forest stands with variable terrain and tree species combinations that represent more extensively forests in Finland than our rather homogeneous test site. The distribution of ALS metrics among different fertility classes also needs investigation. From a practical point of view, it is most critical to detect areas of severe defoliation and test the method with all possible forest site combinations represented. However, it is difficult to predict where and when the mass outbreaks of defoliators will appear. The optimal ALS metric selection method to use in estimating and mapping needle loss also requires further study.