Ability of Remote Sensing Systems to Detect Bark Beetle Spots in the Southeastern US

Research Highlights: Sentinel-2 Normalized Difference Vegetation Index (NDVI) products show greater potential to detect indications of disturbance by bark beetles in the southeastern US than Moderate Resolution Imaging Spectroradiometer (MODIS), as the high spatiotemporal heterogeneity of the southeastern forest land prevents its deployment at the current resolution. Background and Objectives: Remote sensing technologies have been an essential tool to detect forest disturbances caused by insect pests through spectral trait variation. In the US, coordinated efforts such as ForWarn, led by the US Forest Service and based on MODIS satellite data, are used to monitor biotic and abiotic disturbances. Because of the particular characteristics of the southeastern US landscape, including forest fragmentation and rapid forest turnover due to management, detection and visualization of small bark beetle spots using remote sensing technology developed for more homogeneous landscapes has been challenging. Here, we assess the ability of MODIS and Sentinel-2 time-series vegetation index data products to detect bark beetle spots in the Florida Panhandle. Materials and Methods: We compared ForWarn’s detection ability (lower resolution images) with that of Sentinel-2 (higher resolution images) using bark beetle spots confirmed by aerial surveys and ground checks by the Florida Forest Service. Results: MODIS and Sentinel-2 can detect damage produced by bark beetles in the southeastern US, but MODIS detection via NDVI change exhibits a high degree of false negatives (30%). Sentinel-2 NDVI products show greater potential for identifying indications of disturbance by bark beetles than MODIS change maps, with Sentinel-2 capturing negative changes in NDVI for all spots. Conclusions: Our research shows that for practical bark beetle detection via remote sensing, higher spatial and temporal resolution will be needed.


Introduction
Tree mortality has significant effects on the ecology and value of both natural and commercial forests. In the southeastern region of the United States (US), including northern Florida, large-scale tree mortality can severely affect the forestry industry, with both economic and social impacts [1,2]. Monitoring programs that measure disturbances in forest ecosystems have gained importance due to Europe [18]. Here, we comparatively assess the ability of MODIS and Sentinel-2 time-series vegetation index data products to detect bark beetle spots in the Florida Panhandle. Our goal was to compare multiple contemporary approaches for the detection of existing, ground-truthed bark beetle spots to equip managers with information about which method(s) may provide the most feasible avenue for improving bark beetle monitoring tools in the Southeast. We compare ForWarn (MODIS) and Sentinel-2 resolutions and detections in forest stands containing bark beetle spots confirmed by aerial surveys and ground checks from Florida Forest Service. Our objectives are to (a) evaluate the ability to detect spots in MODIS and Sentinel-2 images, and (b) quantify the agreement in NDVI departure between MODIS and Sentinel-2 images.

Experimental Design and Study Area
The impact of Hurricane Michael in October 2018 triggered numerous incipient bark beetle infestations in loblolly (Pinus taeda L.), slash (P. elliottii Engelm.), and sand pine (P. clausa (Chapm ex Engel.) Vasey ex Sarg.) stands in northern Florida. These new outbreaks in previously uninfested forest stands provided us with an opportunity to evaluate the effectiveness of two remote sensing technologies for detection of emerging bark beetle spots in the challenging southeastern landscape: (1) MODIS NDVI change (analogous to the system currently used by US Forest Service's ForWarn disturbance detection tool) and (2) Sentinel-2 NDVI change [10,19]. In order to focus our comparison on these two technologies' abilities to detect bark beetle spots, we used ground-checked spots from the Florida Forest Service's summer 2019 aerial surveys as our study area.
Aerial surveys throughout Northeastern Florida were conducted by the Florida Forest Service from July to August 2019. Potential bark beetle spots in the red-stages, detected through the aerial survey, were georeferenced and ground checked for confirmation during Fall 2019. Only confirmed spots were included to avoid false positives and focus our analysis on the incidence of false negatives, evaluating whether utilizing remotely sensed data from MODIS or Sentinel-2 would lead to omission of active bark beetle spots. When active bark beetle spots were confirmed, we delimited the entire forest stand as a georeferenced polygon. We used true color composites from Sentinel-2 in the July 2019 imagery to estimate the impacted area within each forest stand by delimiting the extent of red tree canopies as georeferenced polygons in ArcGIS. The bark beetle spots occupied a variable area of our forest stand polygons, ranging from~31% to~89%. Therefore, our NDVI measurements for July 2019 were calibrated by the presence of infested and non-infested trees in the studied stands.
Only stands with active spots and active bark beetle galleries that started in 2019 and indicated detectable change in the satellite imagery were selected for the analysis. This time frame was selected based on a bark beetle trapping survey conducted by one of the authors during June 2019. The survey suggested high activity of Ips beetles during June, with observed signs of infestation and spots during August 2019, two months after the high Ips beetle population activity. Active spots without available or cloud-free satellite images matching the dates of the aerial observations were discarded. Stands with tree damage unrelated to beetles, designated as such by the ground survey, were discarded. The confirmed active bark beetle locations included in our study are shown in Figure 1, with county, stand area, estimated infested area within each stand, and survey flight date noted in Table 1.

Image Sources and Baseline NDVI Determination
Sequences of true color images of bark beetle spot locations, obtained from the ESA-Copernicus Sentinel-2 database, were used to determine dates and area of red-brown discoloration of the tree crowns within the forest stands. Visual inspection of these images indicated detectable change in July 2019. Due to high cloudiness in the region, images were inspected to discard spots or time frames with cloud coverage. MODIS 8-day and Sentinel-2 NDVI data products were obtained from Climate Engine to calculate baseline NDVI and create change maps [20]. MODIS 8-day data is gathered twice a day for most areas in the US at a 250-m spatial resolution and processed to 8-day composites. Following a method analogous to that of ForWarn, we first determined a baseline condition of healthy forest for July 2019 by averaging maximum NDVI values for the month of July in the preceding years (2015-2018) [10]. Constructing a multi-year average as a baseline provides a reasonable estimate of the expected NDVI value in the current year in the absence of disturbance. Additionally, the stands selected did not show crown discoloration during the selected baseline period, nor were they directly affected by Hurricane Michael, allowing us to infer that our baseline calculation of NDVI was indicative of healthy forest conditions. Change maps were created by calculating the difference between the NDVI in the month of July 2019 and the baseline NDVI.

Image Processing and Evaluation Methods
Data visualization and analysis were conducted in ArcMap [21]. Images and raster data were clipped to the forest stand polygon shapes. MODIS images were converted to 10 m resolution to allow raster calculations for compatibility with Sentinel-2 data. Departures of NDVI from the "healthy forest" baseline were classified based on magnitude and direction to aid in visualization and interpretation and to allow more refined comparison in the disturbance sensing capabilities of MODIS and Sentinel-2. As even small negative changes in NDVI can indicate insect attack, we considered any negative change in NDVI from the baseline to signify detection of the known bark beetle activity [10,22]. To quantify the strength of correlation between MODIS and Sentinel-2 images, we calculated the Jaccard coefficient [23,24]. This coefficient is calculated as the intersection between the two rasters divided by the union between the two rasters and quantifies similarity. Accuracy of the different satellite images and the disagreement between them is quantified and discussed.

Results
A total of 917.4 ha (2267 ac) with bark beetle damage was observed and ground checked in the Florida Forest Service aerial surveys. Active bark beetle spots in the red-stages were confirmed to have active galleries of Ips species by ground checks. No southern pine beetle spots were found, and no black turpentine beetles (Dendroctonus terebrans) were recorded in the spots studied. After discarding spots with cloud coverage, seventeen stands with active bark beetle spots, corresponding to 479.39 hectares (1185 ac), were used to evaluate the NDVI products from MODIS and Sentinel-2. Within the studied stands, approximately 310.48 ha (767 ac) was infested by bark beetles while 168.90 ha was not infested. From the seventeen stands, Sentinel-2 NDVI image products were able to detect bark beetle spots in all stands. However, MODIS images failed to detect spots in 30% of the stands with known bark beetle infestations. The total number of px identified as harboring a confirmed attack was higher for Sentinel-2 than for MODIS, with 17 ha. (1694 px) undetected by MODIS. Within the studied stands, 343 ha. (847 ac) actively attacked by bark beetles was detected by Sentinel-2, whereas 327 ha. (808 ac) attacked by bark beetles was detected by MODIS. In addition, 152.35 ha of non-infested area was detected by Sentinel-2 and 135.44 ha by MODIS, with a respective difference of 16.55 ha and 33.46 ha from the true color images.
Difference in image resolution between MODIS and Sentinel-2 can be observed in Figure 2. Sentinel-2 NDVI products were able to detect smaller bark beetle spots. Five categories of NDVI departure, based on magnitude and direction, were selected for visualization ( Figure 2, Table 2). Classes 2-4 of NDVI departure were detected by both Sentinel-2 and MODIS. Sentinel-2 captured 71.75% of the negatively departed areas (classes 2-4), while MODIS captured 68.21% (Table 2). However, class 5 of departure, which indicates the largest negative changes detected, was only detected by Sentinel-2. For the five classes analyzed, 49% of px (23,327 px) disagree between Sentinel-2 and MODIS images ( Table 2).
We calculated the strength of agreement between the MODIS and Sentinel-2 NDVI products at the pixel level using the Jaccard coefficient. Sentinel-2 and MODIS images show 75% similarity in detecting any negative change in NDVI (Table 3). However, the agreement of detection of each class differs between satellites, with only 37% similarity for category 2, 12% similarity for category 3, and 0% similarity for categories 4 and 5.

Discussion
Our results show that higher resolution imagery from Sentinel-2 shows greater potential in detecting Southeastern bark beetle spots than MODIS. Sentinel-2 detected negative departures in NDVI for all of our study stands with known bark beetle infestations, while MODIS failed to detect negative changes in NDVI for 30% of the stands. Our study showed a difference between the infested/non-infested area delimited in true color images and the Sentinel-2 and MODIS products, with more infested area detected by NDVI change. Moreover, the extent of detection of each class of NDVI departure differs between satellites. High resolution remote sensing data such as Sentinel-2, Landsat, and WorldView-2 have been used to record and estimate disturbances resulting from bark beetles in Europe and western North America [25][26][27][28][29]. In the eastern US, high-resolution imagery from Landsat 7 was used to develop hazard models for southern pine beetle [30]. In addition, preliminary results by Ritger et al. [19] studying Ips bark beetle outbreaks suggested that higher resolution imagery from Landsat 8 and Sentinel-2 improve precision in locating small bark beetle spots. We focused narrowly on comparing the utility of Sentinel-2 and MODIS products for confirmed active bark beetle infestations to show the potential for improving remote sensing of bark beetle spots. Although known stressed or attacked trees can be detected with hyperspectral remote sensing data, such data are not yet ready for operational use or as a replacement for in situ surveys in the fragmented southeastern US landscape. Similar results were found by Götz et al. (2020). Field verification is still an essential part of the process, adding value to remote sensing images used for forest health assessment [31].
Besides insufficient image resolution, there are also limitations resulting from the nature of bark beetle attacks and the ecological conditions in the Southeast. One of the main ones is that similar spectral trait variation observed for bark beetle attacks is produced by various other types of disturbances [8]. Refining methods for reliably distinguishing these various disturbances will be critical. The solution may lie in greater temporal resolution. Bark beetle spots develop with a certain speed that is greater than most abiotic stresses, while slower than management interventions such as harvest.
The issue of detection timing is particularly difficult in the southeast. Many studies, including ours, are limited to comparing annual index changes resulting from a single time point in a year. That has proven adequate for areas like Europe or northern North America, where pests are limited to one or a few generations per year [32]. However, southeastern bark beetles are active throughout much of the year, potentially having upwards of more than six generations annually, meaning outbreaks can grow quickly over shorter timescales especially in warmer environments [33]. Hurricane and storm events, common to Florida, may increase Ips activity throughout the year due to the increased abundance of breeding material such as fallen branches or boles caused by strong winds. Our area of study was impacted by Hurricane Michael in 2018, triggering numerous incipient bark beetle infestations.
The cloudiness of the region makes it difficult to get reliably clear imagery for the same geographic area over shorter intervals. Satellite imagery from MODIS offers twice-daily available data, providing time series on forest conditions that overcome cloud-free coverage. On the other hand, satellite imagery with lower temporal resolution such as 16-day Landsat and 5-day Sentinel-2, could increase cloud Forests 2020, 11, 1167 8 of 10 coverage by chance. Hirschmugl et al. [34] combined radar data from Sentinel-1 with optical data from Sentinel-2 and Landsat 8 to overcome cloud-cover challenges in tropical forests, and a similar approach may offer opportunities to detect southeastern disturbances more quickly.
The current limitations of remote sensing technologies mean that while we should strive to improve its utility, we should also continue to explore additional tools that can assist managers. Spatial modeling built on ecological understanding of bark beetles may successfully predict and detect early spots. Recent studies on the effect on climatic variables in the activity of bark beetles in the southeastern US, coupled with available information on historical pest occurrence, should be incorporated in monitoring strategies to improve early detection and prediction of spots [35][36][37][38].

Conclusions
Damage produced by bark beetles in the southeastern US can be detected by comparing NDVI products of high-resolution satellites but with a high degree of false negatives, and with spatial resolution exceeding the scope of individual beetle spots. Change detection of any type of remote sensing alone does not explain the cause of the change. Therefore, interpretation after change detection is required to avoid false positives or false negatives. Sentinel-2 NDVI products show greater potential for identifying indications of disturbance by bark beetles than MODIS change maps. Our research shows that using annual index changes with higher spatial as well as temporal resolution may provide a useful tool for forest managers for early bark beetle detection. Continued evaluation of these methods and integration with predictive modeling and improved risk assessment methodology are needed to improve early detection systems such as ForWarn.