Currently, pest insects are the principal biotic drivers causing disturbances threatening Mediterranean forests in combination with abiotic factors such as drought, fire and climate change [1
]. The pine processionary moth (Thaumetopoea pityocampa
Dennis and Schiff.), one of the major defoliating insects in Mediterranean pine forests, has been considered by the Intergovernmental Panel on Climate Change (IPCC) as an indicator of global warming [3
] for being recorded in an expanded biogeographical range of host distribution from Southern Europe in the Mediterranean region towards northern latitudes and higher elevations over the past twenty years [4
]. Consequently, the defoliators may continue to increase with the future trend of climate change scenarios [5
Outbreaks of T. pityocampa
have been observed to be cycles of 6 years on average [2
], mainly in managed young stands [5
]. In a normal year healthy forest stands recover from defoliation, however, periodic outbreaks attack large forests in which host trees may suffer up to 100% defoliation. Thus, severely affected stands may result in significant reductions of individual tree growth, stand productivity, and forest ecosystem health, from tree to stand and landscape levels [5
]. Such growth reduction may accelerate the rate of tree mortality when the damage is accumulated due to other biotic factors as well as abiotic factors [9
]. As a consequence, defoliation focused on host pine trees is more likely to change the structure and species composition in natural stands as well as lose economic values in planted stands [2
]. Furthermore, forest ecosystems at the landscape level are major carbon sinks contributing to mitigate the impacts of climate change [10
]. Therefore, to predict future scenarios on forest productivity and mortality in cases where the outbreaks may occur, reliable data from monitoring the pest distribution at regional and national scales are required [5
Despite the annual ground survey on forest health in the Mediterranean countries such as France, Portugal and Spain which are frequently affected by outbreaks of T. pityocampa
, we currently lack accurate, fine-grained and timely information systems for monitoring the latest forest condition. Thus, an efficient tool to improve such regional or national monitoring systems spatially and temporally may be developed by adequately using the latest remote sensing technologies. In addition to conventional field surveys which are often time-consuming to cover large areas, satellite-based images such as MODIS and Landsat at medium spatial resolution (30–250 m) have been widely used for detecting defoliations in forest systems over the past two decades [1
]. Currently, due to the free accessibility to medium-resolution (10–60 m) products from Landsat 8 [13
] and Sentinel-2 [14
], they are most commonly used for cost-effectively monitoring large areas. For sensors at high spatial resolution (<10 m), the private industry continued to launch satellites such as IKONOS, QuickBird, RapidEye and TerraSAR-X [1
]. With further advancements in spaceborne technology, the sensor’s spatial resolution nowadays can be as high as 0.3 m (WorldView-4), and continue to enhance temporal and spectral resolution as well [15
]. In Spain, using various sensors (Airborne Hyperspectral Scanner, Hyperion, ChrisProba, Quickbird, and Landsat), Cabello et al. [16
] have applied vegetation indices for mapping forest damage caused by T. pityocampa
to estimate the Leaf Area Index in affected pine stands, followed by Sangüesa-Barreda et al. [2
] with the combined method of Landsat-derived vegetation indices and dendrochronology for assessing the tree growth reduction affected by outbreaks of T. pityocampa.
Furthermore, the emergence of airborne laser scanning (ALS) characterized by point clouds complements the three-dimensional (3D) structure in addition to the spatial resolution higher than any spaceborne technology [1
]. Using ALS metrics, classification of defoliated Scots pines at the individual tree level was demonstrated by Kantola et al. [18
Since the cost of above high-resolution products remains a limiting factor for small operational areas, the trend of monitoring forest health in recent studies has shifted to the alternative 3D technology based on cost-effective Unmanned Aerial Vehicle (UAV) at high temporal and spatial resolution over the past decade [19
]. While initial studies with UAVs were focused on crop management for agriculture applications, the latest UAV technology has proved to be effective for forestry applications (forest disturbances and diseases, forest cover mapping, tree species identification, and forest inventory measurements) as a sampling tool for acquiring ground-truth data [20
]. To date only a few studies have examined on the classification accuracy of forest insect defoliations using UAV imagery. The classification methods include Random Forest [19
], object-based image analysis (OBIA) [21
], k-Nearest Neighbor [23
], and maximum likelihood [24
]. Moreover, no study has calibrated satellite-based vegetation indices as a predictive indicator of such defoliation with UAV-derived data as ground-truth while Pla et al. [26
] has recently made progress in calibrating some indices specific to fire damage using UAV imagery.
With the objective to quantify forest response to infestation, we developed remote sensing-derived indicators to measure the defoliation levels. In this study, we aim to evaluate the spatiotemporal degree of defoliation during a recent outbreak of T. pityocampa in Mediterranean pine forests by change detection analysis using a combination of satellite and UAV imagery. Our main objectives are: (1) examine regression models between vegetation indices (VI) derived from Landsat imagery and defoliation degrees interpreted by UAV imagery for calibration; (2) map defoliation classes based on the best-fit VI model to assess classification accuracy for validation.
In Figure 3
the resulting X and Y points were plotted and compared among selected VIs. Moreover, the coefficient of determination (R2
) by dVI is summarized in Table 3
, which was statistically analyzed for logistic regression proposed by McFadden [39
]. The goodness of fit was highest = 0.815 with dMSI while it was not improved by normalizing VIs such as dNDMI, dNDVI, and dNBR.
Using the equations in Table 3
, the range of dVI values were determined as threshold limits in classification (Table 4
). Applying values in dMSI to defoliation %, any value higher than −125 is classified as no defoliation (<10%): −125 to −295 (10–35%); −295 to −453 (35–70%); and lower than −453 (>70%). Based on the threshold limits, pixels assigned to four classes of defoliation were mapped to show the severity across the study area (Figure 4
). The blank pixels in no color indicate either non-forested areas or stands dominated by other tree species, which were initially excluded from the analysis.
The overall accuracy of the threshold classification was presented in Table 5
for the four defoliation classes (nil, low, medium, and high) in the 50 selected samples of dMSI referenced to UAV orthomosaic images. Greatest producer’s accuracy representing a measure of omission error was 90% for the nil defoliation class, where nine out of the 10 cells observed as nil were correctly classified by predicted dMSI. On the other hand, the high defoliation class was mapped with greatest user’s accuracy of 86% indicating commission error, which six out the seven cells predicted as high correctly represented the observed class. Overall, it should be noted that the number of cells classified as nil or medium defoliation was overestimated whereas the one classified as low or high defoliation was underestimated. Finally, the overall accuracy of the classification was 72% calculated by the ratio between the sum of the cells correctly classified from each class and 50 cells in total.
Among five VIs tested in our study, the results with dMSI as the best predictor were consistent with recent studies on insect defoliations [2
]. The SWIR band calculated in dMSI is known to be a good indicator for the plant moisture content in addition to the plant stress detected in the NIR band [1
]. It may be assumed that the MSI better indicated early symptoms of the host trees stressed by dehydration in our study. The first study examining MSI in relation to conifer damage was conducted by Vogelmann et al. [44
], resulted with an R2
of 0.830 in linear regression. Other studies with MSI have been conducted in pine forests by Sangüesa-Barreda et al. [2
] demonstrating the highest significance on ANOVA tests and most recently by Zhu et al. [38
] with an R2
of 0.982 in logistic regression. The MSI was also effectively applied to defoliation in deciduous forests by Townsend et al. [36
] with an R2
of 0.844 in logistic regression and Rullán-Silva et al. [37
] with an R2
of 0.632 in sigmoidal mixed-effects models. For monitoring coniferous forests in general, the MSI has been found to be more effective than the NDVI which has been mainly applied to deciduous forests [38
]. Nevertheless, several potentially robust VIs should be tested on each particular study area since tree-insect relationships vary from site to site [1
Our initial attempt was to use sketch map polygons from field data as training samples for supervised classification on the severity of defoliation. However, the significant discrepancy in spatial resolution between the field data provided by regional rural agents and Landsat data became evident. The sketch map polygons were delineated for classifying severity levels at a coarse scale in hectares including non-forested areas whereas the spatial resolution of Landsat imagery is as fine as 30 m per pixel, which resulted in a wide range of values among pixels within the same polygon. Yet, without any ground observation such as nests of T. pityocampa
, it is often difficult to distinguish the cause of defoliation based on only spectral bands or even aerial surveys at low levels of defoliation [8
]. Further integration by training the rural agents to apply UAV workflows to their annual health survey may fill this monitoring gap. As suggested in the most recent review on forest health monitoring by Hall et al. [35
], how spaceborne and airborne remote sensing may be integrated with aerial and field surveys into a multi-scale, multi-source monitoring system should be explored.
Regarding stand dynamics and species compositions, we extracted pine-dominated stands from the Land Cover Map of Catalonia (MCSC) and assumed that sampled stands were dense enough to represent Landsat VIs based on the defoliation degree of dominant species in pine stands. Nonetheless, we acknowledge the possibility of misrepresenting the VIs in severely defoliated stands to some extent where understory species is non-host evergreen such as Q. ilex
which is not affected by T. pityocampa
. In such cases, healthy understory trees below defoliated pine trees may have reflected more greenness at the stand level. Moreover, it is also possible to overestimate the defoliation degree in stands mixed with non-host deciduous trees shedding their leaves in winter [35
]. This issue of overestimating or underestimating the impacts on host trees can be minimized by discrimination of deciduous species by detecting spectral variations due to vegetation phenology with Landsat time-series approach such as LandTrendr which can monitor cumulative defoliation as well as annual defoliation [45
] while the satellite-based spatial resolution is not high enough to identify individual trees. Thus, for species identification at the tree level recent studies with UAV-derived multispectral bands and the associated indices [20
] may be further investigated to discriminate only those species of interest for calibration and validation of defoliation degrees.
Regression analysis may be improved by increasing the sample size in each severity category or the number of predictive parameters, or testing transformed dVIs [26
]. One simple way to increase the sample size can be achieved by reducing the cell size for sampling VIs from 30 m in Landsat 8 to 20 m in Sentinel-2 imagery. As multiple predictive parameters climate data (temperature and precipitation) or topographic features (elevation, slope and orientation) may be considered to improve the coefficient of determination in regression models. Moreover, we may introduce the number of nests formed by T. pityocampa
as an additional predictive parameter to estimate the infestation severity as recent studies with the UAV technology have attempted to examine the severity of infestation at the individual branch level [24
]. However, the number of nests captured by UAV images from the air may be potentially underestimated if some nests on lower branches or in dense stands are not counted. Regarding observed parameters, our assessment on UAV-derived defoliation levels (%) was limited to manual photointerpretation, including shadows where some uncertainty remains. The automated removal of shadow pixels should be explored in future studies by testing various thresholds on spectral bands.
Threshold classification based on regression models in this study was one method to generate the defoliation severity map with the advantage of increasing or decreasing the number of classes by changing threshold limits of dVI (X) corresponding to the continuous defoliation degree (Y). Other classification methods using non-parametric algorithms may be taken into further consideration with a larger sample size to find the optimal method among unsupervised (ISODATA, K-means), supervised (maximum likelihood), and machine learning such as Random Forest, Decision Trees, k-Nearest Neighbor, and Support Vector Machine [46
]. In general, those studies based on non-parametric models demonstrated that the classification accuracy significantly increased when the number of classes decreased [20
Tradeoffs between spatial, temporal, and spectral resolution are critical to determine classification specific to case study. In our study sketch map polygons served as the primary information to filter affected areas most likely by T. pityocampa
despite low spatial and temporal resolution. The defoliations over winter due to T. pityocampa
can be discriminated by seasonal activity from other potential causes for defoliations during summer such as drought and summer-feeding insects [38
]. Satellite-based Landsat imagery, at medium spatial and temporal resolution, enabled us to calculate various dVIs due to high spectral resolution. While Landsat has the advantage of allowing time-series analysis from the data archive to track back to 1972 [13
], the use of Sentinel-2 has been recently increasing since its launch in 2015 due to the public access available at the higher temporal, spatial and spectral resolution (every 5 days at 10 m, 20 m, or 60 m for 13 bands) [14
] than that of Landsat 8 (every 16 days at 15 m, 30 m, or 100 m for 11 bands). As shown in Figure 4
, those dVIs at medium spatial resolution need to be calibrated with field observations to estimate the severity classification at a regional scale. Thus UAV imagery observed at high spatial and temporal resolution may improve the efficiency of such calibration. Yet, current limitations of UAV technology include battery duration for 20–30 min, associated small area coverage for sample images per flight, and imagery acquisition permission due to privacy issues specific to some countries [20
]. To cover a large area at landscape and regional scales, it would require multiple flights that may not be consistent with the time, sensor and weather conditions, therefore, it would not replace satellite-based imagery.
Nonetheless, the latest UAV can obtain dense point clouds and multispectral bands (sensors for red edge, NIR and SWIR outside the visible spectrum), which may be a promising technology with a high spatial and spectral resolution for small-scale forestry applications. Using the density of points at a tree level, as successfully demonstrated by Näsi et al. [23
], the structural change in individual trees may be detected and monitored for cumulative defoliation. While some studies [23
] have used the UAV-derived NDVI as the most robust indicator for their analysis on insect defoliations, future studies shall compare it to the UAV-derived MSI which can be calculated from SWIR in the latest sensor technology. Where high spectral resolution is not required for small operational areas, compared to using both satellite and UAV imagery, preparing a UAV flight would greatly increase the time-efficiency and cost-effectiveness as well as flexibility in planning imagery acquisition [19
]. In addition to such advantages as alternative methods for generating orthomosaic images and calculating VIs from multispectral bands, the use of UAVs enables to avoid clouds during flights, which often cannot be controlled by satellite orbit scheduling [19
]. Thus, with relatively less efforts and lower costs, UAV imagery may increase the spatial quality to be equivalent to ground-truth data.