Utility of commercial high-resolution satellite imagery for monitoring general flowering in Sarawak, Borneo
Tomoaki Miura and Yuji Tokumoto contributed equally to this study.
Abstract
General flowering (GF), irregular synchronous mass flowering of multiple tree species across multiple families, is a unique biological phenomenon of the mixed lowland dipterocarp forest in Southeast Asia. Characterizing the spatial extent and temporal dynamics of GF is essential for an improved understanding of climate–vegetation interactions and the potential climate change impact on this species-rich rainforest. We investigated the utility of newly available high-temporal (daily) and high-spatial (3–4 m) resolution remote sensing by the PlanetScope commercial satellite constellation for detecting flowering trees in a dipterocarp rainforest at Lambir Hills National Park, Sarawak, Malaysia. Our study was focused on the latest GF event known to have occurred in the region in the year 2019. PlanetScope successfully acquired 13 clear-sky or minimally cloud-contaminated scenes over the park during a study period of January 1, 2019 to August 31, 2019 encompassing the 2019 GF event. In situ phenology observations verified that the PlanetScope images detected the flowering crowns of tree species that turned into white or orange. This multitemporal image dataset also captured the flowering peak and species differences. The correlation coefficients between the multitemporal image signatures and in situ phenology observations were moderate to very strong (0.52–0.85). The results indicated that the 2019 GF event was a whole-park phenomenon with the flowering peak in May. This study reports the first successful satellite-based observations of a GF event and suggests the possibility of regional-scale characterization of species-level phenology in the dipterocarp forest, key information for biodiversity conservation in Southeast Asia.
1 INTRODUCTION
The lowland mixed dipterocarp forest once covered extensively the lowland areas of West Malesia and is one of the most biologically diverse forests in the world (Myers et al., 2000). This forest encompasses three of the terrestrial ecoregions of the world, each of which is considered a distinct assemblage of natural communities and species and should be monitored for biodiversity conservation (Olson et al., 2001). A recent global multicriteria analysis also ranked most of these lowland areas at high priority for ecosystem restoration for biodiversity conservation (Strassburg et al., 2020).
A unique and, probably, the most striking feature of this climax lowland tropical rainforest is the phenomenon of “general flowering (GF)” (Appanah, 1985; Ashton et al., 1988). At irregular intervals often after one to 10 years of sterility, a large number of tree species from different families flower synchronously for a few weeks to a few months, which is then followed by mast fruiting (Appanah, 1985; Ashton et al., 1988; Sakai, 2002). Trees in various plant taxa likely flower at one period to enhance the pollination success and disperse the fruit, seed, and/or seedling predation pressures (Sakai et al., 2006). At the same time, trees that share a similar complement of pollinators seem to flower sequentially in order to reduce interspecies competition for pollinators (Appanah, 1985). The region over which such a mass-flowering event occurs can be as small as a single river valley, or as large as northeastern Borneo or peninsular Malaysia (Ashton et al., 1988; Numata et al., 2003).
Both local plot-based and regional-scale studies suggested drought and cool temperature as the key triggers of GF (Azmy et al., 2016; Kobayashi et al., 2013; Numata et al., 2013; Sakai et al., 2006; Yamasaki et al., 2017). Unlike the temperate region that exhibits clear annual cycles in climate variables, the climate of this region is aseasonal and lacks a regular annual dry season (Inoue et al., 1993). Recent molecular phenology and modeling studies showed that drought and cool temperature not independently but synergistically drove GF (Y.-Y. Chen et al., 2018; Ushio et al., 2020; Yeoh et al., 2017). The requirement for both cool temperature and drought for floral transition implied that GF should be sensitive to climate change (Yeoh et al., 2017). Y.-Y. Chen et al. (2018) also found that drought and cool temperature thresholds varied among species. Thus, the capability to observe and monitor the temporal and spatial dynamics of the individual crown-level or species-level flowering phenology at a regional scale is critical to understand the potential impact of climate change on the lowland mixed dipterocarp forest for their biodiversity conservation (Ichie et al., 2004).
Previous GF observations have been limited either in the spatial or temporal extent. In nearly all previous studies, the timing and magnitude of GF events have been observed by in situ field methods in defined plots or along roads at weekly, bi-weekly, or monthly intervals (Chechina & Hamann, 2019; Y.-Y. Chen et al., 2018; Numata et al., 2003; Ushio et al., 2020). Nagai et al. (2016) employed a daily time-lapse camera to obtain species-level phenological traits, including flowering, in this species-rich dipterocarp forest, but with a limited spatial coverage and, thus, a limited sample size per species.
Satellite remote sensing has been shown to be an effective means for regional-scale phenology observations (Dixon et al., 2021; Landmann et al., 2018). Phenological remote sensing of dipterocarp forests of Southeast Asia has, however, continued to pose a challenge due to persistent cloud cover in the region. While medium resolution (10–30 m) satellite data such as Landsat and Sentinel-2 have the potential to observe individual crown-level phenology, their utility is limited as they often fail to image land surface due to their low temporal resolution (5- to 16-day repeat cycle) (Torbick et al., 2017). Coarse resolution (250 m to 1 km) satellite data such as Moderate Resolution Imaging Spectroradiometer (MODIS) have temporal resolution high enough (1- to 2-day repeat cycle) to obtain cloud-free scenes, but their spatial resolution is too coarse to obtain phenology information in this biologically-diverse forest. For example, MODIS green-red vegetation index time series data could not detect surface color changes due to flowering in this region (Nagai et al., 2014).
The advent of high-temporal and high-spatial resolution commercial satellite remote sensing has brought about the possibility of observing and characterizing GF in unprecedented spatial detail (Brandt et al., 2020; Wagner et al., 2020). PlanetScope, in particular, can obtain high spatial resolution (3–4 m) images on a nearly daily basis (Planet Team, 2017; Roy et al., 2021) and has been applied to a wide range of ecological and environmental studies, especially for detecting ephemeral and/or small changes on landscapes (Cheng et al., 2020; Cooley et al., 2019; Csillik et al., 2019; Helman et al., 2018; Xu et al., 2020; Yamano et al., 2020). Sakuma and Yamano (2020), for example, used PlanetScope time series data to classify crop types on a subtropical island in Japan. Likewise, Wang et al. (2020) demonstrated the utility of PlanetScope time series data to assess seasonal changes in both ecosystem- and crown-scale spectral reflectance in Central Amazon.
Remote sensing studies of flowering plants have been based on the spectral reflectance changes of the plants associated with the flowering (Landmann et al., 2018). Landmann et al. (2018) summarized with their extensive review that flowering generally resulted in an increase in the visible (VIS) reflectance and a decrease in the near-infrared (NIR) reflectance, and that “most of the recent studies relied solely on the VIS region, confirming the suitability of the VIS spectrum.” In recent studies, PlanetScope data captured the bloom dynamics of almond orchards (light pink to white flowers) in the Central Valley of California (B. Chen et al., 2019), delineated the in situ observed flowering season of subalpine meadows (a mixture of various flower colors including white, violet, and purple) in the Cascade Range of Washington (John et al., 2020), and captured the flowering dynamics in eucalypt-dominated forest landscapes (white flowers) in Western Australia (Dixon et al., 2021). In addition, Paz-Kagan et al. (2019) successfully mapped two highly invasive Acacia species using their prominent yellow flowers in a coastal plain in Israel with WorldView-2 satellite imagery, another commercial high-resolution platform.
- Would the daily temporal resolution be high enough to capture the flowering phenology of a GF event in this persistently cloud-covered region?
- Would the 3–4 m pixel size be high enough to track the flowering phenology at individual crown level in this dipterocarp forest?
- What flower colors could PlanetScope multispectral bands detect in this taxonomically diverse forest?
We selected Lambir Hills National Park, Sarawak, Malaysia as our study site as GF in the park has been studied by various research groups for a number of years. Our investigation was focused on the latest 2019 GF event known to have occurred in the park.
2 MATERIALS AND METHODS
2.1 Site description
Lambir Hills National Park (4°12′N, 114°2′E) is ca. 7000 ha in area and is located close to the center of the northwestern coast of the Malaysian part of Borneo, the third-largest island in the world (Figure 1). The southwestern section of the park is underlain by the Mid-Late Miocene Lambir Formation (sandstones and shales), whereas the younger Pliocene Tukau Formation (sandstones) underlies the northeastern part of the park (Kessler & Jong, 2015; Liechti et al., 1960). The soils of the park are derived from these sandstones and shales (Lee, Davies, et al., 2002). Most of the areas are covered by lowland dipterocarp forest dominated by trees in Dipterocarpaceae, Anacardiaceae, and Sterculiaceae in the upper layer, and those in Euphorbiaceae, Myristicaceae, and Burseraceae in the lower layer. Kerangus forest, a tropical heath forest, covers the high-elevation areas on the western side of the park (Figure 1).
In 2001, an 80 m canopy tower crane was constructed and a 4-ha permanent forest research plot (200 m × 200 m; 150–250 m above sea level, ASL) centered at the crane was established (Ozanne et al., 2003) (“LBR (Lambir) Tower” in Figure 1). This 4-ha plot is located in the east of a 52-ha permanent plot (Lee, Ashton, et al., 2002). Mean annual temperature and rainfall are 26°C and 2600 mm, respectively, based on data recorded at the plot from January 2001 to December 2009 (Kume et al., 2011). A tree census is conducted every year (Nakagawa et al., 2012). Diameter at breast height (DBH, diameter of trees at 1.3 m above ground level) is measured for trees over 10 cm DBH, and all trees with DBH >10 cm are identified into genus and/or species.
Previous field observational studies reported nine GF events during a 27-year period from 1993 through 2019 in the park: 1996, 1997, 1998, 2001–2002, 2004, 2005, 2009–2010, 2013–2014, and 2019 (Iku et al., 2017; Sakai, 2002; Sakai et al., 1999; Sakai et al., 2006; Ushio et al., 2020).
2.2 PlanetScope data
PlanetScope is a constellation of approximately 130 satellites in low-Earth (450–580 km altitude), Sun-synchronous orbit (Planet Team, 2021). The constellation is able to acquire near-nadir images of the entire Earth's land surface on a daily basis. Each satellite is of a 3U CubeSat form (10 cm × 10 cm × 30 cm) and acquires multiband images at 3.7–4.1 m spatial resolution with 12-bit radiometric resolution. Three generations of PlanetScope satellites are currently in orbit with the first, second, and third generations referred to as Dove or PS2, Dove-R or PS2.SD, and SuperDove or PSB.SD, respectively. The Dove/PS2 satellite has four spectral bands located at the blue (455–515 nm), green (500–590 nm), red (590–670 nm), and NIR (780–860 nm) wavelength regions.
All PlanetScope images acquired over the study area between January 1, 2019 and August 31, 2019, a period encompassing the 2019 GF event, were examined for cloud contamination using the Planet Explorer online image viewer (https://www.planet.com/explorer). Those images not subject to significant cloud cover were marked and obtained from the Planet Explorer. We selected orthorectified, surface reflectance products (Planet-designated Level 3B) which were projected onto the Universal Transverse Mercator coordinate system based on the World Geodetic System 1984 datum with 3-m grid spacing. Positional accuracy of the orthorectified products is less than 10 m (root mean square error) (Planet Team, 2021). The atmospheric correction was made with the 6SV radiative transfer code (Vermote et al., 2006) constrained with the spatially and temporally closest MODIS aerosol optical thickness data (Planet Team, 2020).
As reported in previous studies, the acquired PlanetScope surface reflectance scenes were subject to slight temporal radiometric inconsistency (Csillik et al., 2019; Leach et al., 2019; Wegmueller et al., 2021). We located two targets that appeared unchanged during the study period (a cemetery and a grassland area) in the south of Lambir Hills National Park. Pixel values over the two targets were extracted from every PlanetScope image. Using these pixel values, each band of every scene was radiometrically cross-calibrated to the image acquired on June 18, 2019. We assumed a simple linear model in deriving these cross-calibration factors.
To examine the advantage of PlanetScope's daily temporal resolution, we assessed the availability of near cloud-free image data from other high-to-medium spatial and lower temporal resolution multispectral satellite sensors, including WorldView-2 and -3, Sentinel-2, and Landsat-7 and -8. WorldView-2 and -3 are commercial high-resolution (1–2 m) data sources and acquire images upon user's requests. Thus, the temporal resolution of available WorldView-2 and -3 data is highly variable. Sentinel-2 acquires 10-m resolution images in the VIS–NIR regions every 5 days. Landsat-7 and -8 acquire 30-m resolution images every 16 days. As their repeat cycles are 8-day apart, a combination of Landsat-7 and -8 provides an 8-day temporal resolution.
2.3 Field observations
Twelve tree individuals located in the 4-ha plot were observed in situ for flowering phenology with a pair of binoculars from the top of the crane. These trees included five species: Dryobalanops aromatica, Shorea beccariana, Shorea ochracea, Swintonia foxworthyi, and Pentace borneensis (Table 1). They were emergent trees or in the canopy layer, composing the upper layer of the forest. The reproductive organs of these species were white or yellow (Table 1), whereas their leaves were green on the adaxial side, and light green or greenish-yellow on the abaxial side. In addition, two unidentified Shorea trees were located, one in the 4-ha plot and the other adjacent to the 4-ha plot. Although we were unable to identify their exact reproductive stages, these two trees drastically changed their crown color from green to orange. For each of the 14 trees, the percentage of the tree crown in flower (or in orange color for the two Shorea spp.) was recorded approximately bi-weekly from February 18, 2019 to June 19, 2019, except for the second half of May 2019 in which no in situ observations were made. All 14 trees were confirmed to be located in the area covered by the satellite images described in the “PlanetScope Data” section above by using the coordinates recorded during the tree census. Thus, they were suitable for comparison with satellite-based detection of flowering trees.
| Family name | Species name | Observed number of tree individuals | Tree ID used | Observed phenological event | Observed canopy color change |
|---|---|---|---|---|---|
| Dipterocarpaceae | Dryobalanops aromatica | 4 | DA1–4 | Flowering | Green to white |
| Shorea beccariana | 5 | SB1–5 | Flowering | Green to pale yellow | |
| Shorea ochracea | 1 | SO1 | Flowering | Green to whitish yellow | |
| Shorea sp. | 2 | SS1–2 | Unknown | Green to orange | |
| Anacardiaseae | Swintonia foxworthyi | 1 | SF1 | Flowering | Green to white |
| Tiliaceae | Pentace borneensis | 1 | PB1 | Flowering | Green to white |
Digital photographs of the forest cover in and around the plot, including the 14 tree individuals, were taken from the top of the crane as visual documentation of the in situ phenology observations with the binoculars. The photos were taken with an iPhone 8 (Apple Inc., USA) or Ricoh WG-5 (Ricoh Imaging Company, Ltd., Japan) hand-held camera with an auto-adjustment mode in brightness and white balance.
2.4 Time-lapse digital imagery from the Phenological Eyes Network
Two in situ time-lapse digital image datasets of the canopy layer around the 4-ha plot were obtained from the Phenological Eyes Network (PEN) to temporally interpolate the in situ binocular phenology observations (Nagai et al., 2018). One dataset was acquired with an automatic digital fisheye camera (ADFC) system made with a Coolpix 4500 digital camera with a FC-E8 fisheye lens (Nikon, Tokyo, Japan), and an SPC31A controller (Hayasaka Rikoh, Sapporo, Japan), whereas the other was acquired with a camera module attached to a Raspberry Pi B+/3 computer (Raspberry Pi Foundation, Cambridge, United Kingdom). The ADFC system was attached near the top of the crane at 76 m above ground level (AGL), facing westwards. The Raspberry Pi system was installed on the crane at 64 m AGL facing southwards. The ADFC system took digital images at the local noon time on a daily basis, whereas the Raspberry Pi system took digital images hourly between 6:00 and 18:00 local time every day. The Raspberry PI system's real-time clock (RTC) module was found to gradually drift. The RTC time was reset on February 18, 2019 and was approximately 50 min earlier than the actual time when it was checked for maintenance on September 17, 2019.
Six tree individuals (DA1–4, SB2, and SO1) out of the 14 were within the field of view of either of the 2 PEN digital cameras and their crown color changes were observable in the PEN time-lapse image dataset. Using the binocular observations and handheld camera photos as a reference, the PEN time-lapse images were visually inspected and the flowering phenology was estimated for each of the six trees: the start of flowering that corresponded to the date of the crown color change from green, the peak flowering period that corresponded to the highest flowering intensity observed by the binoculars, and the end of flowering that corresponded to the date when the crown color changed back to green.
2.5 Statistical analysis
We performed correlative analysis between PlanetScope image pixel data and in situ flowering phenology observations for the six tree individuals for which continuous phenology observations were made from the analysis of PEN time-lapse images (DA1–4, SB2, and SO1). The PEN-based flowering phenology data were converted into numerical values. The value of “one” was assigned to the peak flowering period, whereas the value of “zero” was assigned to nonflowering dates. Numerical values were assigned to the transition dates by assuming a linear increase from zero to one for the period between the start of flowering and the start of the peak flowering period and a linear decrease from one to zero for the end of the peak flowering period to the end of flowering. Spearman correlation coefficients were computed for each of the six trees.
3 RESULTS
3.1 In situ phenology observations
GF was observed in situ from March 2019 through July 2019 in and around the 4-ha research plot. In particular, the canopy color of various plant species turned to white in May (Figure 2). Figure 3 graphically summarizes in situ observed flowering dates and/or periods of the 14 tree individuals.
The binocular observations accompanied with the handheld digital camera photos found that the 12 tree individuals flowered in three groups in sequence (filled circles in Figure 3; see Figure S1 for representative pictures of flowering trees): the flowering of S. beccariana (SB1–5) on April 12–14, 2019; D. aromatica (DA1–4), S. forxworthyi (SF1), and P. borneensis (PB1) on May 3–5, 2019; and S. ochracea (SO1) on June 13–19, 2019. The five S. beccariana trees (SB1–5) changed their crown color gradually from green (February 20, 2019) to pale yellow (March 27, 2019, April 14, 2019, and May 05, 2019). The four D. aromatica trees (DA1–4) had the color changes from green (March 9, 2019 and April 14, 2019) to white (May 3, 2019) and back to green (June 15, 2019). The S. foxworthyi (SF1) crown color changed from green (March 27, 2019) to white (May 3, 2019) to dull yellow in June (June 15, 2019). The S. ochracea tree crown (SO1) was green in April and May 2019, and became yellowish in June 2019. Unidentified Shorea trees (SS1 and SS2) turned from green to orange in June 2019 (Table 1).
The overall period of the crown color change was about 40–50 days based on the PEN time-lapse camera dataset (Figure 3; see Figures S2–S4 for sample PEN time-lapse camera images representing flowering phenology). The peak flowering period was ca. 10 days for D. aromatica (DA1-4), ca. 20 days for S. beccariana (SB2), and ca. 30 days for S. ochracea (SO1). For S. ochracea (SO1), we could not identify when flowering ended because of the large bracteoles of the species, but the crown color turned reddish around July 19 and thereafter.
3.2 PlanetScope imagery
Clear, partially cloud covered, or thin-cloud covered scenes of PlanetScope Dove/PS2 satellites were available for 13 days for the 8-month study period (Table 2). All these scenes, except for the January 4, 2019 image, were acquired between 10:00 a.m. and 10:30 a.m. local time with comparable sun-target-sensor geometries, that is, nadir viewing with the Sun at 30° to 40° zenith angle in the North–East–East direction. The January 4, 2019 image was acquired with the Sun in the south direction.
| Image acquisition date | View zenith angle (°) | Sun zenith angle (°) | Sun azimuth angle (°) |
|---|---|---|---|
| January 4, 2019 | 1 | 41 | 133 |
| April 13, 2019 | 2 | 32 | 80 |
| April 17, 2019 | 3 | 30 | 76 |
| May 2, 2019 | 1 | 30 | 67 |
| May 5, 2019 | 4 | 34 | 67 |
| May 8, 2019 | 1 | 30 | 63 |
| May 12, 2019 | 2 | 31 | 61 |
| May 22, 2019 | 5 | 36 | 60 |
| May 27, 2019 | 3.5 | 34 | 58 |
| June 12, 2019 | 0 | 34 | 56 |
| June 18, 2019 | 5 | 39 | 57 |
| July 9, 2019 | 3 | 35 | 56 |
| August 18, 2019 | 3.3 | 31 | 71 |
WorldView-2 and -3 acquired no image over the study area during the 8-month study period. Sentintel-2 had near-cloud-free scenes on the three dates of January 4, 2019, June 18, 2019, and July 13, 2019, but all images acquired in April and May 2019, during which a large number of trees flowered, were cloud-contaminated. A combination of Landsat-7 and -8 similarly had near-cloud-free images on three dates (March 6, 2019, June 18, 2019, and August 21, 2019), but all images in April and May 2019 were subject to considerable cloud cover.
Figures 4 and 5 are PlanetScope natural color composite images of Lambir Hills National Park for three dates on or around which binocular field observations confirmed the flowering of trees (see Figure 3). Whereas a limited number of white dot features were found in the April 17, 2019 image (Figures 4a and 5a–c), a large number of clumped white dot features were visible in the May 5, 2019 image (Figures 4b and 5d–f). When compared with surface elevation data (Figure 1), these white dot features were observed from 50 m to 300 m in elevation across the entire park and varied in size and density (Figure 5d–f). The same clumped white dot features were clearly visible in the May 2, 2019 and May 8, 2019 images, and the locations and shapes of the features remained the same through the three images. Thus, those white dot features likely corresponded to the crowns of flowering trees.
In the June 18, 2019 image, a small number of not only white, but also reddish white dot features were seen (around the center and in the upper half of Figure 5i). Reddish white dot features were also visible in the July 9, 2019 and August 18, 2019 images, and their locations remained the same across the images when found on all three images or any two consecutive images. Thus, those reddish white dot features likely corresponded to the crowns of flowering trees or those in some distinct phenological stage as well.
In Figure 6, PlanetScope natural color composite images over the 4-ha research plot are shown for the same three dates as in Figures 4 and 5, and also for July 9, 2019, along with the crown locations of the 14 tree individuals. The image pixel colors corresponding to the SB1–5 locations were slightly brighter on April 17, 2019, but did not appear yellow and remained nearly the same throughout the four dates. The image pixels corresponding to the locations of DA1–4, PB1, and SF1 were white-colored on May 5, 2019, but were not distinguishable from the surrounding on the other dates. The image pixels for SO1 were brighter and light green to white-colored on June 18, 2019, and reddish on July 9, 2019. Last, the image pixels corresponding to SS1 and SS2 were reddish white on June 18, 2019 and July 9, 2019. All these color changes corresponded well to the timing of flowering (or crown color change for SS1 and SS2) observed in the in situ datasets (Figure 3). These results indicate that PlanetScope captured flowering for white and orange, but not yellow.
Image pixel (reflectance) values (2-pixel-by-2-pixel window) corresponding to the locations of the 14 tree individuals were extracted from the 13 PlanetScope images (Figure 7). They were also extracted from a location to the south of the 4-ha plot in which no color change was observed (thus, referred to as “unflowered” trees hereafter). For the unflowered trees, blue, green, and red reflectances remained unchanged throughout the four dates, whereas the NIR reflectance increased slightly (Figure 7a). In contrast, all VIS reflectances (blue, green, and red) were higher on May 5, 2019 when flowering was observed for D. aromatica (DA1), S. foxworthyi (SF1), and P. borneensis (PB1) (Figure 7b–d). NIR reflectance changed only slightly for D. aromatica (DA1) (Figure 7b), but decreased after flowering for S. foxworthyi (SF1) and P. borneensis (PB1) (Figure 7c,d). For S. beccariana (SB5), VIS reflectances varied only slightly, whereas NIR reflectance decreased 1 month after the flowering period ended (Figure 7e). The VIS reflectances of S. ochracea (SO1) and Shorea sp. (SS2) increased when flowering was observed (Figure 7f,g). For all those trees for which flowering was visible on the images (D. aromatica [DA1], S. foxworthyi [SF1], and P. borneensis [PB1]), NIR reflectance increased slightly at the peak of flowering, but then was lower than the preflowering period after flowering (Figure 7b–d).
Given the VIS reflectance increase associated with flowering, we computed a simple ratio vegetation index (SR) of the red reflectance divided by the NIR reflectance to enhance the spectral reflectance change associated with flowering. The flowering crowns were depicted well as having higher SR values (Figure 8). The SR was also higher for the crane.
SR values extracted from the 13 PlanetScope images are plotted for the 14 observation trees and unflowered trees (Figure 9). For the unflowered trees, the SR remained nearly unchanged for the study period (Figure 9a). For D. aromatica, S. foxworthyi, and P. borneensis, the timing when the SR increased corresponded well to the peak flowering period (Figure 9b,c). For S. beccariana the SR did not increase during the peak flowering period (April 2019), but was slightly lower in June and July (Figure 9d). For S. ochracea, the SR was the highest in July when the crown appeared reddish white, and decreased in August when the crown became reddish in PEN image data (Figures 9e and S3). For the two Shorea spp. (SS1 and SS2), the SR peaked in June and July, respectively, when the field data indicated that the crown turned orange and also when the crowns appeared reddish in the satellite image (Figure 9e).
In Table 3, the derived correlation coefficients between the SR and PEN-based flowering phenology are summarized for the six tree individuals. Except for DA1, the sample sizes were less than 13 because the crane was right above the crowns of those trees in some of the 13 images, decreasing the number of available image data days. The correlation coefficients were high for D. aromatica, being 0.61 or higher and significant at 5% or higher level (Table 3). The correlation was moderate at 0.52, but significant at 5% level for S. ochracea. In contrast, the correlation was weak at 0.33 for S. beccariana. The statistical analysis results corresponded well with the visual analysis of the SR temporal profiles.
| Species name | Tree ID | Sample size (n) | rs |
|---|---|---|---|
| Dryobalanops aromatica | DA1 | 13 | 0.65** |
| DA2 | 11 | 0.85*** | |
| DA3 | 11 | 0.79* | |
| DA4 | 11 | 0.61* | |
| Shorea beccariana | SB2 | 8 | 0.33 |
| Shorea ochracea | SO1 | 12 | 0.52* |
- *, **, and *** denote significant at 0.05, 0.01, and 0.001, respectively.
4 DISCUSSION
This study evaluated the potential of PlanetScope high-spatial (3–4 m) and high-temporal (daily) resolution satellite data to detect flowering crowns and track their flowering phenology during the 2019 GF event known to have occurred in the lowland mixed dipterocarp forest of Lambir Hills National Park, Sarawak, Malaysia. With the daily temporal resolution, the PlanetScope satellite constellation successfully acquired 13 cloud-free or little cloud-contaminated scenes during the 8-month study period of January 1, 2019 to August 31, 2019. Flowering trees were observed as white or reddish white dot features in the 3–4 m satellite pixels, which were verified with in situ phenology observations made in the 4-ha plot located within the park. This PlanetScope multitemporal dataset depicted the flowering phenology of select tree species and distinguished the differences in the timing of flowering peak among tree species. The analysis results indicated that the 2019 GF event at the park was a whole-park phenomenon, the flowering occurring for the elevation range of 50 m to 300 m throughout the lowland dipterocarp forest, and that the flowering peak occurred in early May 2019. This study demonstrated the utility of this new satellite remote sensing technology for characterizing the spatial extent of GF and the flowering phenology associated with GF.
4.1 Flower colors
In situ phenology observations conducted near the crane verified that three flowering tree species from different families all having white-color flowers (i.e., D. aromatica; S. foxworthyi; and P. borneensis) were observed with PlanetScope spectral bands. They also verified that the satellite data detected a species with whitish-bright yellow flowers (S. ochracea) and the orange-colored crowns of unidentified Shorea species. Genera of our target species, Dryobalanops (Dipterocarpaceae) and Swintonia (Anacardiaceae) have the reproductive organs on terminal shoot or basal axilla with numerous flowers (i.e., paniculate or raceme inflorescences) (Ashton, 2004; Kochummen, 1996) and flowerings of these taxonomic groups are clearly visible from the crane due to the drastic crown color changes. The results of this study indicate that other species with white flowers in the same genus should also be detectable from the satellite images. Indeed, one Swintonia species, Swintonia acuta that has white flowers on raceme, changes the crown color from green to white in flowering events and has been detected by time-lapse cameras at the same site (Nagai et al., 2016).
The flowering crowns of S. beccariana, which bears numerous pale-yellow flowers on the terminal shoots, were not seen in the satellite images. Leaves of S. beccariana are green on the adaxial side, but yellow on the abaxial side. It is likely that PlanetScope's spatial resolution of 3–4 m was simply not high enough to capture flowering crowns of S. beccariana, given the smaller contrast in color between S. beccariana's pale yellow flowers and the yellow leaf colors on the abaxial side. The flowering of S. beccariana was even difficult to see in in situ handheld and PEN time-lapse camera images without trained eyes. Nevertheless, tree species with white color flowers could potentially serve as indicator species for satellite-based detection of a possible GF event where they exist in abundance.
4.2 Temporal resolution of satellite imagery
PlanetScope's daily temporal resolution increased the number of cloud-free image acquisitions during the 2019 GF event, that is, 13 images over a 8-month period (ca. 240 days). This was approximately a 5% chance of obtaining a cloud-free image. The Sentinel-2 platform and the Landsat-7 and -8 constellation each acquired three cloud-free images over the same period, that is, approximately a 1.2% chance of obtaining a cloud-free image. Our results were comparable to those found in the conterminous United States for which Moon et al. (2021) found that PlanetScope imagery had a higher frequency of clear-sky acquisitions (1.0- to 3.9-day intervals) than a 4-day resolution harmonized Landsat-8 and Sentinel-2 image dataset (4.1- to 11.0-day intervals).
The 13 cloud-free images were not equally distributed within the 8-month period, but concentrated in April–June with six images in May. This 3-month period coincided with the flowering period of the tree species examined in this study, allowing our PlanetScope multitemporal dataset to depict and compare the flowering phenology of select tree species. The same analysis should be conducted at a large number of locations to further examine the appropriateness of the daily temporal resolution for capturing the flowering phenology of a GF event.
Whereas the daily temporal resolution of PlanetScope improved the chance of obtaining cloud-free scenes, clouds still presented a problem as some scenes, while useful, contained patchy and/or thin clouds. In order to further improve the utility of PlanetScope or high-resolution satellite data for GF studies, an operational algorithm to correct optically-thin clouds is desirable. Flowering canopies are visible and recognizable under optically thin clouds; however, correction would allow for quantitative analysis of the features. It would also be possible to increase the temporal resolution by having another set of satellites with a different overpass time to image the same areas more than once per day, given the high diurnal variations of cloud cover in the region.
4.3 Perspective for future phenology monitoring
Previous studies have shown the importance of species-level phenology information and continuous in situ measurement data of species traits (e.g., leaf photosynthetic capacity and leaf area index) for developing community- or ecosystem-level vegetation phenology models (Muraoka et al., 2010; Tang et al., 2016). The results of this study also indicate the importance of continuous in situ measurements of individual tree-level crown traits from both the biological and remote sensing perspectives. The traits of interest include flower colors, fruit colors, changes in crown colors and reflectance associated with flowering and fruiting, and crown structures (where flowers and fruits are borne). Continuous in situ measurements should help determine remotely sensible plant phenological traits that are of biological and ecological importance. This should, in turn, allow us to better interpret high-resolution satellite signals and to regionally scale up and interpolate significant phenological traits with the satellite images. The “catalogued” plant phenological trait information should also help develop a good understanding of the species discrimination capability with high-resolution remote sensing. Different tree species should be discriminable with high-resolution satellite images only when their remotely sensible phenological traits for that satellite are different. Multiscale data acquisitions that involve data from field-based trait measurements (Sakai, 2002) and molecular expression data (Kobayashi et al., 2013; Yamasaki et al., 2017; Yeoh et al., 2017), to time-lapse digital imagery (Alberton et al., 2014; Lopes et al., 2016; Nagai et al., 2016) and drone imagery (Dixon et al., 2021; Fawcett et al., 2020; Klosterman et al., 2018), should be a promising approach for developing such a catalogue as also discussed by Tang et al. (2016) and Dixon et al. (2021).
Regional-scale detection of flowering phenology should contribute significantly to deepen our understanding of GF triggers. When they are used to extend point-based field measurements, satellite remote sensing can provide the flowering phenology information in a spatially-continuous manner for much larger areas. Such data can serve as input to regional-scale studies of GFs and improve the analysis (Azmy et al., 2016; Numata et al., 2013). In Numata et al. (2013), for example, spatial variations of GFs could not be associated with meteorological factors well due to coarse and mismatching resolutions of in situ GF observations and meteorological factors. Azmy et al. (2016) used satellite-derived rainfall factors to pair meteorological factors with in situ GF observations more accurately. Both types of analysis can greatly benefit from regional-scale satellite-derived flowering phenology data.
ACKNOWLEDGMENTS
This study was conducted under a Memorandum of Understanding between the Sarawak Forest Department (SFD, Kuching, Sarawak, Malaysia) and the Japan Research Consortium for Tropical Forests in Sarawak (JRCTS, Japan). The authors thank the Forest Department Sarawak and all members of the Lambir Hills National Park site community. The PlanetScope data used in this study were made available through the Planet's Education and Research program (Tomoaki Miura). This study was partially supported by a NASA grant (80NSSC19K1691), MEXT KAKENHI Grant Numbers 18H04785 and 22H05179, JSPS KAKENHI Grant Number 19K15875, University of Zurich University Research Priority Program (URPP) Global Change and Biodiversity (GCB), Beatrice Ederer-Weber Foundation, Georges and Antoine Claraz-Donation 2019, MEXT LEADER (2020L0080), the Global Change Observation Mission (PI #117, ER2GCF104) of Japan Aerospace Exploration Agency (JAXA), and a JST CREST grant (number JPMJCR16O3).
CONFLICT OF INTEREST
The authors declare no conflict of interests.
