Comparison of piece length and dynamics of large wood in streams covered with coniferous and broadleaf forests mapped using ortho-photographs acquired by an unmanned aerial vehicle

As large wood (LW) supplied by landslides and debris ows is one of the main components of watershed ecosystems, the importance of quantifying the dimensions of the LW is evident. However, the low accessibility of disturbed channels after landslides and debris ows generally impedes accurate and quick LW investigations. Recent advances in photogrammetry techniques may overcome such issues. In this study, we used ortho-photographs acquired using a small unmanned aerial vehicle (UAV) to measure the lengths of LW (wood pieces > 0.2-m long and > 0.03-m diameter) entrapped mainly by closed-type check-dams. We focused on two channels that are located in coniferous and broadleaf forests and affected by two different landslides events. The measurement accuracy was analyzed by comparing the lengths derived from the UAV method with direct measurements. When the both ends of a piece LW are satisfactorily extracted from an ortho-photograph acquired via the UAV, the LW lengths of coniferous trees can be measured with an accuracy of approximately ±0.5 m. For broadleaf trees, most of the extracted lengths were shorter than the directly measured lengths, probably due to the low visibility arising from the complex structures of the root wad and tree crown. Most LW pieces were discharged from landslide scars in the broadleaf forest, whereas approximately 750 LW pieces were left in the landslide scars of the coniferous forest. The number of LW pieces in the landslide scars increased with the increase in the landslide area, suggesting that some LW pieces can be left even if large landslides occur. There were no signicant changes in the lengths or locations of the entrapped LW, at either site seven months after the rst UAV ight. In the coniferous forests, the rainfall that triggered landslides in 2017 exceeded the 100-year return level, which was an abnormally intense rainfall. Although the 2019 rainfall event that occurred between UAV ights did not provide enough rainfall to trigger landslides, rainfall intensities with different durations reached the second-highest value from 1976 to 2019, exceeding the 30-year return period. This suggests that most of the entrapped LW rarely migrate even under extreme rainfall.


Introduction
Debris ow causes the entrainment of stand-woods located in initiation and riparian zones, and consequently it may include ~10%-60% of uvial large wood (LW) by volume (Johnson et al., 2000;May and Gresswell, 2003a;Lancaster et al., 2003). In addition to its destructive impact on life and infrastructure (e.g., Ruiz-Villanueva et al., 2013), wood pieces in a debris ow can alter the ow regime because of their irregular shapes, which can be entrapped around obstacles and lead to anomalous deposition of sediment and inundations (e.g., May, 2002;Lancaster et al., 2003;Tang et al., 2018;Booth et al., 2020). The accumulation of LW results in structural peculiarities around channel networks (e.g., Keller and Swanson, 1979;Woodsmith and Bu ngton, 1996;Montgomery et al., 1996;Nakamura and Swanson, 2003). This contributes to changes in the ecosystem, the channel morphology, and the sediment ux through logjam formation and the decay of LW pieces (e.g., Wallace and Benke, 1984;Lisle, 1995;Montgomery et al., 1995;Gurnell et al., 2001;Comiti et al., 2006;Ruiz-Villanueva et al., 2016).
Hence, the river form and function are determined by the interaction between water, sediment, and wood (Nakamura et al., 2017;Swanson et al., 2020). Therefore, quantifying LW is important for assessing its impacts on ecological, geomorphological, and uvial conditions. Many previous studies focusing on LW pieces in streams have contributed to unraveling the role of LW in various spatiotemporal-scales and environmental settings, but most of these approaches required direct eld measurements, such as local monitoring (e.g., Manners et al., 2007), the tracking of LW pieces (e.g., Ravazzolo et al., 2015;Wyżga et al., 2017), and eld experiments using arti cial wood pieces (e.g., Haga et al., 2002). Taking into account the diversity of forests around channels (e.g., age, species, and density of trees), the monitoring and eld survey of LW pieces is evidently an effective approach. However, the necessity of human effort in the eld hinders data acquisition in inaccessible areas (e.g., headwater channels and disturbed areas immediately after landslides and debris ows). In practice, eld-data acquisition is di cult over large-scale areas exceeding the size of a sub-basin. The improvement of eldmeasurement techniques for LW thus remains a critical issue.
To address the measurement issues for LW, remote sensing approaches using three-dimensional data have been applied. The Use of LiDAR (light detection and ranging) data clearly reduces the processing time required for mapping logjams and LW (Kasprak et al., 2012;Abalharth et al., 2015;Atha and Dietrich, 2016). However, this technology is expensive and is therefore available only in some regions.
Alternatively, photogrammetry based on structure-from-motion multi-view stereo (SfM-MVS) using a UAV (unmanned aerial vehicle) can produce three-dimensional data (e.g., digital-surface-models and pointclouds), which has been proven to be time-e cient compared to classical eld surveys (Sanhueza et al., 2019). This approach overcomes data-availability issues and is relatively low cost. Nevertheless, most tests have been conducted in lowlands and ood plains rather than in low-accessibility areas such as steep channels (e.g., Sanhueza et al., 2019). As the accuracy of three-dimensional data acquired by SfM-MVS is greatly in uenced by complex surfaces and obstacles-such as steep slopes, large reliefs, and vegetation coverage (e.g., Fonstad et al., 2013;James and Robson, 2014)-many unresolved uncertainties remain over the application of the SfM-MVS approach using three-dimensional data for LW measurements in steep and complex areas such as channels impacted by landslides and debris ows.
Aerial photography is a traditional source of two-dimensional data that may provide more-or-less meaningful information about LW. Even satellite images from Google Earth are being used for accurate LW mapping (Atha, 2013;Ulloa et al., 2015). Hence, the e cacy of aerial photography is evident, but the accuracy and effort of mapping depend on the image quality. In this respect, it is expected that small UAVs may allow the acquisition of high-resolution aerial photographs at low cost because of the lower ight altitudes and higher portability of UAVs compared to conventional aerial vehicles. Moreover, because ights of small UAVs can overcome inaccessibility issues and cover several kilometers, depending on the ight design, a UAV is a fairly attractive tool for obtaining LW measurements in lowaccessibility areas.
In addition, LW is often quickly removed by river administrators to avoid unexpected damages in the downstream area resulting from its transport by stream ow after being supplied and entrapped in streams (hereinafter this is referred to as secondary transport). The risk arising from the secondary transport of LW (e.g., damages on facilities in residential regions) has not been evaluated properly so far, because it has been di cult to conduct eld surveys of the LW supply immediately after intense rainfall.
In contrast, UAV ights can be conducted immediately after rainfall events that result in a large amount of LW, which can potentially contribute to unraveling LW dynamics after their supply and entrapment. Therefore, LW mapping can enable investigations of LW transport due to subsequent rainfall, even if it is carried out with a simple method using aerial photography. Nevertheless, the potential to measure LW based on aerial photographs acquired via UAV has not been thoroughly examined due to lack of sample cases.
In this study, we analyze the accuracy of UAV-based measurements for LW lengths. Two regions are selected representing two forest types, coniferous and broadleaf forests. In both regions, large amounts of LW were supplied to streams through landslides and debris ows triggered by a single rainfall event. This research has two main objectives: (1) to analyze the capability of ortho-photographs acquired via UAVs to measure the lengths of LW, and (2) to investigate whether entrapped LW pieces move again due to intense rainfall. Based on the results, we discuss the effectiveness of UAV measurements and how LW behave after their entrapment.

Study Sites
To examine the in uence of tree type on the accuracy of LW measurements obtained using the orthophotographs acquired by the UAV, we selected two different sites covered by coniferous or broadleaf forests ( Figure 1a). Hereafter, the former site is referred to as the CF (coniferous forest) site (Figure 1b), while the latter site is referred to as the BF (broadleaf forest) site ( Figure 1c). In this study, we use the term "tree type" to indicate either coniferous or broadleaf trees rather than the detailed tree species.

Coniferous forest site
The CF site is a sub-watershed of the Otoishi watershed located in the northern part of Kyusyu Island, Japan (Figure 1a In the region around the CF site, an intense rainfall on July 5, 2017, triggered more than 2,000 landslides, which resulted in debris-ow propagations and therefore a great amount of LW (Chigira et al., 2018). The sliding sediment layers were mainly granodiorite and pelitic schist (Chigira et al., 2018). The intense rainfall in July 2017 triggered six landslides at the CF site ( Figure 2a), but the existing check-dams effectively prevented the descent of sediment and LW. Hence, most of the LW pieces were trapped around the check-dams and along the channel, thereby preventing damage in the downstream residential areas. Because of this low impact on the residences, the LW pieces were not removed and were left in-situ until the time we carried out observations.

Broadleaf forest site
The BF site is located in the Hiroshima prefecture in the western part of Japan ( Figure 1c). Here, two channels are con uent at an irrigation pond that is located at the lower ends of channels (~305 m a.s.l., Figure 2b). The highest points of channels A and B are ~410 and ~480 m a.s.l., respectively. Their total lengths are ~880 and ~466 m, respectively. The average slopes of the two channels are similar, at ~11°-12° (before the debris-ow occurrence). The drainage areas of channels A and B are ~0.07 and ~0.13 km 2 , respectively. Similar to the case of the CF site, four closed-type check-dams were constructed before the debris ow occurred in 2018 ( Figure 2b). The forest type is completely different, and most of the hillslopes are covered by broadleaf forests with various tree types. The tree heights range from approximately 10 to 20 m.
In the Hiroshima prefecture, a stationary front affected by Typhoon Prapiroon caused heavy rainfall on July 5-7, 2018 (Tsuguti et al., 2019) and triggered approximately 8,000 landslides due to the vulnerable geological setting, which is mainly dominated by weathered granite (Kaibori et al., 2018). Five landslides occurred at the BF site, but the check-dams and the pond effectively prevented the descent of sediment and LW. Similar to the situation at the CF site after the disaster, the LW pieces produced in the BF site were left in-situ until we carried out the observations.

Methods
This study consists of two parts: (1) an assessment of the accuracy of the LW measurements obtained from ortho-photography acquired via the UAV, and (2) a rainfall analysis to investigate the possible transport of entrapped LW.

UAV ights and aerial-photograph processing
A small consumer-grade UAV (DJI Mavic 2 pro; Table 1) was used for the ights covering the channels and landslides ( Figure 2). Two UAV ights were conducted at each of the respective study sites (Table 2): (1) February 14 and September 26, 2019, at the CF site, and (2) November 16, 2018, and December 25, 2019, at the BF site. The ights were conducted manually at elevations between approximately 50 and 200 m, and most of the photographs were taken in the nadir direction. The ight path provided at least a 70% overlap of the aerial-photographs. To reduce the differences in shade and brightness among the acquired aerial-photographs, the ights were operated to avoid strong sunlight.
The acquired images were processed using SfM-MVS photogrammetry software (Agisoft, Metashape Professional version 1.5.1). Usually, arti cial reference markers are established, and their absolute coordinates are measured with a GNSS (global navigation satellite system), providing ground-control points (GCPs) for geo-referencing the SfM-MVS photogrammetry (e.g., Barnhart et al., 2019;Tsunetaka et al., 2020). However, immediately after a debris-ow disaster, the establishment of arti cial reference markers is impeded by LW pieces, which behave as obstacles. The recently developed, high-accuracy, RTK (real time kinematic) GNSS-based direct geo-referencing system built into SfM-MVS photogrammetry processing makes it possible to obtain accurate absolute coordinates without GCPs (e.g., Carbonneau and Dietrich, 2017). However, the consumer-grade UAV we used does not include such an RTK-GNSS system, because it is an expensive technology. Because of these operational limitations, we focused on the ability to make LW measurements using relatively low-cost UAV-based processing. Our intention was thus to assess the accuracy of measurements using only UAV ights without any other eld operations (e.g., setting arti cial reference-markers) or RTK-GNSS systems.
The position data from the built-in GNSS that controlled the ight position were contained in exchangeable image le format data for each photograph. Using these data, we preliminary collected the coordinates of the processed ortho-photographs. Following standard SfM-MVS photogrammetry work ow, the ortho-photographs were produced with a spatial resolution of 0.03-0.05 m. Because the measurements were performed without establishing arti cial reference markers, the coordinates of the resulting ortho-photographs taken during the different ight dates did not overlap accurately. As the principal aim of using sequential ortho-photographs was to investigate the transport of the LW, it is necessary only to have relative coordinate system overlap (i.e., absolute coordinate system overlap is not required). Thus, we manually adjusted two sequential ortho-photographs by using the crests of the checkdams as GCPs for which the coordinates were taken from the older ortho-photograph. Using the extracted coordinates, the ortho-photograph acquired on the later date was transformed to t the rst one. The differences in the lengths of selected immobilized objects (e.g., large boulders and the crests of checkdams) were less than approximately 0.2 m. Therefore, we considered the horizontal accuracy of the ortho-photographs to be less than 0.2 m.

Accuracy assessment
Because we used ortho-photographs (i.e., two-dimensional data) to quantify the LW, the dimension of LW measured by the UAV is its length. First, we extracted the length of the LW, both as a single piece and as an element of a jam-forming log as line data from the acquired ortho-photographs. Note that invisible LW pieces below a jam-forming log are out of the measurement target. All visible LW pieces were extracted at the CF site (the area surrounded by the dotted white line in Figure 2a), whereas only LW pieces around the check-dams A and B were investigated at the BF site because of the low visibility of LW originating from broadleaf trees (Figure 2b). Accordingly, the measurement range differed from a reach scale at the BF site to a sub-watershed scale at the CF site.
The measurement accuracy of the extracted length is presumably affected by visibility, which is related to the types and sizes of the trees. Moreover, because LW is often inclined and broadleaf trees have a curved shape, orthogonal projection to obtain the extracted length may cause a measurement error ( Figure 3).
Hence, we focused on the differences among the tree types (i.e., coniferous or broadleaf trees), diameters, and entrapment inclinations of the LW.
To compare with the extracted lengths of the LW, we directly measured LW pieces at the study sites ( Figure 3). Based on the ortho-photographs from the rst ight, the visible LW pieces were labeled. When we conducted the second ight, we found and directly measured some of the labeled LW pieces (50 LW pieces at the CF site and 131 LW pieces at the BF site). In the found LW pieces, we measured the lengths -excluding the root-wads and twigs-directly using a tape (estimated precision < 0.05 m), similar to the approach of Sanhueza et al. (2019) and compared those measurements with the lengths extracted from the ortho-photographs from the rst UAV ight (Figure 3). Using the same tape, we measured the circumferences of the found LW pieces at approximately 1.0 to 1.5 m height from the root-wad. For LW pieces less than approximately 1.5 m in length, because there were only slight differences in the thickness of a single fragment, we measured the circumference at an arbitrary location. From the measured circumferences, we calculated the diameters of the found LW pieces. Note that, because the minimum measurable circumference was approximately 0.1 m, some of the calculated diameters of the LW pieces were smaller than the estimated precision of the tape measurement (0.05 m). At the CF site, as the LW retained relatively straight shapes, the entrapment inclination was measured directly using an angle meter (estimated precision < ±1°, Figure 3a).
The extraction of LW depends on its visibility in the ortho-photographs, and thus the identi cation of its origin as well as the detection of small wood pieces of less than the minimum resolution of the orthophotographs (i.e., 0.03 m in diameter) were beyond the measurement purpose. In addition, the successful extraction of LW pieces from the ortho-photographs was constrained by the visibility of the LW length, not by the visibility of the LW width (diameter). Considering the horizontal accuracy of the ortho-photographs (< 0.2 m), the minimum length of the extracted LW pieces was set to 0.2 m. Although LW is typically de ned as a wood piece > 0.1 m in diameter and > 1 m in length (e.g., Jackson and Sturm, 2002;Tang et al., 2018), we de ned LW as wood pieces > 0.03 m in diameter and > 0.2 m in length in this study, following the minimum measured diameters and lengths of the LW. Moreover, the extracted LW may contain multiple pieces that originated from a single tree. Note that, as a single stand-wood may result in several pieces of LW, counting the extracted LW involves uncertainty.
Analysis to determine the transport of large wood To analyze whether or not LW migrated through time, we created the quantile-quantile (Q-Q) plots of the probability density of the extracted length. The CF site was divided into seven zones: six landslide scars and the riparian zone with a channel approximately 760 m in length (Figure 2a). Changes in the probability density of the LW length were visually tested using the Q-Q plots. Similarly, Q-Q plots were made for channels A and B at the BF site (Figure 2b). For each zone, to investigate the coincidence of the probability density of the extracted length between the two ight dates ( In addition, we investigated whether speci c rainfall characteristics triggered the secondary transport of LW. For this reason, we used the hourly rainfall records from 1976 to 2019 obtained at the closest raingauges installed by the Japan Meteorological Agency, the Asakura station (for the CF site, Figure 1b) and the Higashihiroshima station (for the BF site, Figure 1c). The annual maxima of rainfall intensity of various durations (1, 2, 3, 6, 12, 24, 48, and 72 hours) were investigated from 1976 to 2019 and compared with the intensity of the rainfall that triggered the debris ows. In addition, we examined the records of rainfall maxima after the debris-ow occurrence. The return periods were calculated based on the probability density of each rainfall intensity tted by a Gumbel distribution based on the annual maxima (e.g., Koutsoyiannis et al., 1998;Sane et al., 2018). For this tting process, the goodness-of-t test (Kolmogorov-Smirnov test) resulted in p-values exceeding 0.1, suggesting that the applied Gumbeldistribution models well tted the probability density of the observed rainfall.

Results
Lengths of the entrapped large wood As expected, a comparison of the extracted and measured LW lengths indicates that the measurement accuracy depends on the visibility of the LW. Fifty pieces of LW were directly measured at the CF site ( Figure 4a). Fourteen of them were trapped by stand-woods located in the riparian zone, and they were partially invisible from the sky due to the coverage by the tree crowns. This low visibility caused underestimates of the extracted lengths, with errors in length being a maximum at ~10 m. In contrast, 36 other pieces were fully visible, and their lengths ranged between about 3 and 22 m. Despite this wide range, most measurements coincided well with the extracted lengths. This clear dependence on the visibility of the LW is re ected in the boxplot ( Figure 5), indicating that the interquartile ranges obviously differed between the visible and covered LW; the differences ranged from −0.4 to 0.5 m and from 1.6 to 4.3 m, respectively ( Figure 5).
The lengths of 130 LW pieces without crown coverage were directly measured at the BF site, but they rarely coincided with the extracted lengths, which were usually underestimates ( Figure 4b). Moreover, the differences between the lengths varied. The related interquartile range varied between approximately 0.3 and 2.1 m, re ecting the lower accuracy and higher uncertainty when measuring broadleaf trees than when measuring coniferous trees ( Figure 5).
Strictly speaking, the extracted length was the orthogonally projected length, and we therefore calculated the slope distance of the measured LW from the extracted length using the measured slope. However, this correction did not signi cantly improve the measurement accuracy because differences in the visibility of the LW depending on the difference in the tree types overwhelmed the in uence arising from the approximation by the orthogonal projection (Figures 4a and 6). The differences in the diameters of the LW indicate that the extracted lengths of relatively narrow pieces (<~0.2-0.3 m in diameter) at the BF site tend to be underestimated (Figure 7), despite the absence of crown and logjam coverages. Our de nition of LW differs from the usual de nition of LW as wood pieces with diameters exceeding 0.1 m (e.g., Jackson and Sturm, 2002;Tang et al., 2018). Note that, especially for LW pieces with diameters less than 0.1 m, the direct measurements (< 0.05 m) and the ortho-photographs (< 0.2 m) can contain signi cant measurement error independent of the differences in tree types. However, even for LW pieces with diameters between 0.1 and 0.2 m, the measurement errors were higher at the BF site than at the CF site (without coverage, see Figure 7). This suggests that the decrease in the visibility of small-diameter pieces was responsible for the decreased measurement accuracy at the BF site.
Most LW pieces were discharged from landslide scars at the BF site, whereas a total of approximately 750 LW pieces (exceeding 50% of the total of ~1,370 LW pieces) were left in the landslide scars at the CF site. At the CF site, the interquartile of the length of the LW differed slightly among the landslide scars, but they were similar between the riparian zone and the average of all landslides: 2.5 to 7.7 m and 2.6 to 8.3 m, respectively (Figure 8a). The median length of the LW ranged from 3.4 m (in landslide 4) to 7.8 m (in landslide 1). At the BF site, the interquartile of the length of the LW at check-dams A and B were 1.1 to 3.1 m and 0.8 to 2.6 m, respectively (Figure 8b). The median lengths of the LW at check-dams A and B were similar: 1.8 m and 1.5 m, respectively. The median length of the LW at the CF site was ~4.5 m, which is evidently greater than that at the BF site (Figure 8c).
The areas of the landslide scars at the CF site ranged widely, from ~800 to 5,400 m 2 (Figure 9). The number of entrapped LW pieces tended to increase with the increase in the area of the landslide scars ( Figure 9). Landslide 2 (~5,400 m 2 ) accounted for approximately 40% of the total area of the landslide scars (i.e., the total landslide area at the CF site is ~14,000 m 2 ), and thus this landslide was the predominant sediment source at the CF site. Despite this huge sediment discharge, the number of entrapped LW pieces was highest among the landslide scars, at approximately 350 LW pieces Transport of entrapped large wood At both sites, there were only slight visible changes in the spatial distribution of the LW between the two ortho-photographs (Figures 10 and 11). The LW pieces were trapped in various areas-for instance, as stand-woods in the riparian zone, the check-dams, and the landslide areas. Despite these differences in the entrapment zones, it seems that most of the LW did not move even after about 7 months (at the CF site, Figure 10) and over one year (at the BF site, Figure 11). The Q-Q plots indicate that the probability density of the extracted LW length coincided well with the measured length, regardless of the division of the entrapment zones ( Figure 12). Thus, the p-values with respect to the goodness-of-t test exceeded 0.05, suggesting that the changes were statistically nonsigni cant.
The rainfall anomalies that triggered debris ows were revealed by comparing the annual maxima of the rainfall intensity in each case. At the CF site, the rainfall that triggered the debris ow in 2017 was obviously excessive rainfall, as the rainfall intensity was the highest compared with the annual maxima of other years, regardless of the rainfall duration (Figure 13a). This suggests that the highly intense and continuous rainfall resulted in the large amount of LW. At the BF site, for the rainfall that triggered debris ows in 2018, rainfall intensities with 1-6 hour durations were not higher than the other annual maxima since 1976, but those with 12-72 hour durations were the most intense since 1976 (Figure 13b). This suggests that a high rainfall intensity continuing over 12 hours resulted in landslides and debris ow accompanied by the propagation of LW. Both rainfall events that triggered debris ows at the CF and BF sites reached levels exceeding the 100-year return period (Tables 3 and 4).
At the BF site, after the landslides and debris ows in 2018, the maximum annual rainfall intensity in 2019 was not remarkable (Figure 13b, Table 4). Thus, although most LW pieces remained in place, it is di cult to analyze the potential of secondary transport of entrapped LW via intense rainfall. In contrast, at the CF site, after the landslides and debris ows in 2017, the annual rainfall intensity maxima in 2018 and 2019 were the second or third highest values since 1976 (Figure 13a). Although the intensity of the triggering rainfall in 2017 signi cantly exceeded the annual maxima in 2018 and 2019 in terms of the return period, some rainfall intensities in 2018 and 2019 exceeded the level of the 30-year return period (Table 3). Such a heavy rainfall event occurred on July 21, 2019 (in between the UAV ights at the CF site; Table 2).

Implications for UAV-based large wood mapping
Our results demonstrate that the visibility and tree types of the LW determine the ability to measure their lengths accurately using ortho-photographs acquired via UAVs. The lengths of most of the LW pieces originating from coniferous trees were measured with better than ±0.5 m accuracy using the orthophotographs acquired by the UAV, but for the broadleaf trees, the lengths were underestimated systematically despite the absence of tree-crown coverage reducing the visibility (Figures 4b and 5).
Considering the complex structures of both ends of the broadleaf trees (i.e., wide and rounded crowns and root-wads; Figure 3b), these underestimates may be attributed to missing extractions due to the low visibility of the broadleaf trees arising from the complex structures at the ends. In addition, the lengths of the entrapped LW at the BF site was less than that at the CF site (Figure 8), implying the fragmentation of broadleaf trees during the debris ow. Although the characteristics of the debris ow also probably in uence the magnitude of LW fragmentation (e.g., Johnson et al., 2000;Lancaster et al., 2003), the fragility of broadleaf trees may reduce the measurement accuracy based on the ortho-photograph at the BF site by decreasing the visibility of the LW.
Hence, most of the factors that decrease the measurement accuracy of LW depend on visibility. Based on this interpretation, LW measurements via UAV have both positive and negative aspects. Several components that affect the quality of the aerial photographs (e.g., black shadows, brightness involving overexposure, and overlap ratios among the aerial photographs) depend on the weather conditions and the ight plan (e.g., the path, height, and direction of shooting). For example, it is expected that low ight altitudes and ights under cloudy conditions can avoid block-out shadows and blown-out highlights in the acquired UAV images, resulting in an increase in the visibility of LW in the processed orthophotographs. Thus, an adequate ight plan may signi cantly improve the accuracy of UAV-based LW measurements, without additional processing in the SfM-MVS photogrammetry.
Conversely, our results reveal an obvious limitation of ortho-photograph-based measurements. In general, LW is not only dispersed along a channel, but also results in logjam due to stacking and accumulation (e.g., Manners et al., 2007;Abalharth et al., 2015). However, the correlation between measurement accuracy and visibility from the sky demonstrates that the length measurements of LW pieces covered by a logjam are beyond the scope of ortho-photograph-based measurements. Sanhueza et al (2019) pointed out that vegetation in streams also may cover LW pieces and thus decrease the measurement accuracy of LW pieces based on images acquired by UAV. Indeed, the tree-crown coverage decreased the measurement accuracy of LW length (Figure 4a). Given these measurement limitations, it is especially di cult to measure jam-forming LW pieces in streams deeply covered by vegetation (e.g., riparian trees) using the approach employing images acquired by a UAV.
In other words, the total amount of LW is probably underestimated in all UAV-based measurements. In this respect, the entrapment of LW pieces in landslide scars at the CF site may also have been underestimated, suggesting that high-ratio LW pieces were left even in slipped hillslopes. In some subwatersheds with coniferous forests located around the CF site, no LW pieces were left in the landslide scars (Marutani et al., 2017). Thus, the difference in tree type (i.e., coniferous or broadleaf trees) cannot account for the high-ratio of LW entrapments in the landslide scars at the CF site compared with that at the BF site. The median length of LW varied among landslide scars (Figure 8a), probably depending on differences in runout processes and sediment volume among landslides. Considering this, the increase in the number of entrapped LW pieces with the increase in the landslide area ( Figure 9) indicates that LW pieces may remain in landslide scars regardless of differences in size and the dynamics of landslides. It seems that the outlets of landslide scars at the CF site are relatively narrow (like bottle-necked shapes) compared with those at the BF site (Figure 2), which may lead to high-ratio LW entrapments.
Possibility of the transport of the entrapped large wood The slight changes in the lengths and spatial distributions of the LW suggest that the LW pieces that were entrapped by the check-dams did not move after the debris-ow disaster (Figures 10, 11, and 12). The elapsed time from the rst ight was about seven months for the CF site and over a year for the BF site (Table 2), which may be insu cient to cause signi cant visible decay. This means that intact LW still remains in the upper reaches of the channels. However, considering the slight differences in the spatial distribution and the Q-Q plot of the LW at the CF site (Figures 10 and 12), the extremely high rainfall intensities in 2019 (Figure 13a and Table 3) suggest that most of the entrapped LW remains in place and rarely migrates, even when intense rainfall exceeding the level of the 30-year return period occurs. This implies that most rainfall events that do not cause debris ows and landslides cannot trigger secondary transport of entrapped LW.
Because the pre-existing check-dams trapped LW at both sites, differences in the trend between such arti cial trapping and natural entrapment of LW should be noted. Even so, the low transportability of the entrapped LW provides insight into the role of the LW and sediment deposited in the upstream channel after debris ows and landslides. The e cacy of LW trapping by closed-type check-dams is generally considered relatively low compared with that of open-type check-dams (e.g., Piton and Recking, 2016).
Nevertheless, the entrapment of LW pieces by closed-type check-dams continued in the long term, due to the low possibility of secondary transport at our study sites. May and Gresswell (2003b) argued that LW pieces in streams can facilitate sediment deposition, thereby resulting in an increase in sediment storage.
Indeed, the low transportability of the entrapped LW pieces in the CF and BF sites indirectly indicates low erosion and discharge rates of the sediment stored after a landslide occurrence, suggesting that the entrapped LW pieces can impede subsequent sediment transport. For this reason, the trapping of LW pieces in low-order streams during and immediately after a landslide occurrence is notably important in mitigating subsequent impacts on facilities in residential regions due to the LW pieces and sediment supplied.

Conclusions
We mapped large wood (LW) pieces using ortho-photographs acquired using a UAV in two headwater channels covered by coniferous or broadleaf forests (the CF and BF sites, respectively) to: (1) determine whether or not ortho-photography acquired via UAV allows measurements of the lengths of the LW, and (2) to investigate the possibility of the transport of entrapped LW by intense rainfall.
At the CF site, comparing the directly measured lengths with the lengths extracted by UAV mapping indicates that aerial-photography measurements had approximately ±0.5 m accuracy when the both ends of the LW were satisfactorily visible. In contrast, some LW pieces were trapped by stand-woods in the riparian zone, which were invisible underneath the tree crowns, and this led to underestimation of the extracted lengths. At the BF site, most of the extracted length values were shorter than the lengths measured directly, probably due to low visibility of the ends of the LW pieces due to complex tree structures in the root-wads and tree crowns. Hence, these results suggest that the measurement accuracy of LW via UAV strongly depends on the visibility of the LW arising from tree types. In other words, although ight operations that adequately increase the visibility of LW in UAV images may improve the accuracy of LW length measurements, the UAV-image-based approach is fundamentally inadequate for measuring jam-forming LW pieces, especially in streams covered by riparian trees.
Most LW pieces did not remain in the landslide scars at the BF site, whereas approximately 750 LW pieces were left in the landslide scars at the CF site. The number of LW pieces in the landslide scar tended to increase, depending on the increase in the landslide area. Therefore, even large landslides cannot transport all LW pieces from their slip hillslopes in some cases, suggesting that LW pieces may remain in landslide scars regardless of the landslide area.
At both sites, changes in the frequency distribution of the lengths and the locations with respect to entrapped LW were slight even over seven months had passed since the rst mapping. Nevertheless, at the CF site, some rainfall intensities between the two ights reached the second-highest values since 1976. Although the rainfall intensity that triggered the landslides in 2017 was higher than the annual maxima in 2019, some intensity values in 2019 exceeded the 30-year return period. This suggests that most of the entrapped LW remains and rarely migrates again, even under intense rainfall. The data used in this paper are available from the authors upon request.    Boxplot of differences between the extracted and measured lengths of the large wood (LW). Differences in length were calculated by subtracting the extracted length from the measured length.