Warren and entrance detections by thermal imager

Thermal imaging technology is a developing field in wildlife management.  Most thermal imaging work in wildlife science has been limited to larger ungulates and surface-dwelling mammals.  Little work has been undertaken on the use of thermal imagers to detect fossorial animals and/or their burrows.  Survey methods such as white-light spotlighting can fail to detect the presence of burrows (and therefore the animals within), particularly in areas where vegetation obscures burrows.  Thermal imagers offer opportunity to detect the radiant heat from these burrows, and therefore the presence of the animal, particularly in vegetated areas.  Thermal imaging technology has become increasingly available through the provision of smaller, more cost-effective units. Their integration with drone technology provides opportunities for researchers and land managers to utilise this technology in their research/management practices.  
We investigated the ability of both consumer (<AUD$20,000) and professional imagers (>AUD$65,000) mounted on drones to detect rabbit burrows (warrens) and entrances in the landscape as compared to visual assessment. 
Thermal imagery and visual inspection detected active rabbit warrens when vegetation was scarce. The presence of vegetation was a significant factor in detecting entrances (P<0.001, α=0.05).  The consumer imager did not detect as many warren entrances as either the professional imager or visual inspection (P=0.009, α=0.05).   Active warren entrances obscured by vegetation could not be accurately identified on exported imagery from the consumer imager and several false-positive detections occurred when reviewing this footage. 
We suggest that the exportable frame rate (Hz)was the key factor in image quality and subsequent false positive detections.  This feature should be considered when selecting imagers and suggest that a minimum export rate of 30Hz is required. Thermal imagers are a useful additional tool to aid in identification of entrances for active warrens and professional imagers detected more warrens and entrances than either consumer imagers or visual inspection.

Methods
We used three uncooled microbolometer arrays (Table 1) of varying sensor size and cost.  The Jenoptik VarioCAMⓇ HD (hereafter referred to as the “Jenoptik”) professional thermal imager was used to evaluate part 1, with the FLIR Zenmuse XT640 and Sierra-Olympic VayuHD used in part 2 (hereafter referred to as the “Zenmuse” and “Vayu” respectively).  The Zenmuse came as an integrated system with the DJI Inspire 1 drone; however, both the Jenoptik and the Vayu were heavier non-integrated imagers.  Both of these imagers required mounting to a Ronin MX gimbal (https://www.dji.com/au/ronin-mx) for image stabilisation.  The Jenoptik was mounted to a DJI S1000+ drone (https://www.dji.com/au/spreading-wings-s1000/spec ) and the Vayu mounted to a DJI Matrice 600 drone (https://www.dji.com/au/matrice600/info#specs).  All video was collected and processed as “white-hot” grayscale imagery.  

Determining which warren entrances belong to which warrens can be challenging in high density rabbit populations.  For the purposes of this research, an entrance was part of the same warren if it was within 5m of another entrance.  When an entrance was detected that was more than 5m away from another entrance, this was deemed to be part of a new warren.  Single entrances that were >5m away from other entrances were considered a single-entrance warren.  Warrens were regarded as active when one or more entrance had signs of use.  This includes a lack of vegetation growing in the entrance, the presence of freshly excavated soil, fresh scat and/or the presence of rabbit footprints.  Warrens where all entrances were covered in either debris (leaves and sticks), with cobwebs and with hard crusted soil were considered inactive.  No further validation (e.g. excavation or trapping) was undertaken to confirm warren activity status.

All thermal imager surveys were conducted in the morning before first light to maximise the temperature differential between warren entrances and the surrounding terrain.  All sites were visually inspected for rabbit warrens (active and inactive) on foot during the day (prior to the thermal survey) and all identified warrens were mapped with their GPS locations recorded.  The ground and aerial surveys were independent, i.e. the thermal imager transects were designed prior to visual inspection.  In Part 1 we determined whether active rabbit warrens could be detected with a thermal imager. We flew the drone with the Jenoptik imager directly to the warren locations. In Part 2 we compared a professional imager (Vayu) to a consumer imager (Zenmuse).  We established parallel flight transects to allow complete coverage of the area being investigated and to mimic the actual survey method that should be employed to search for warrens.  We undertook visual counts of warrens and warren entrances in Part 2.  Visual counts were undertaken upon arrival and before the drone flights.  Parallel line transects approximately 10m apart were walked and all warrens and associated entrances were recorded.  Once imagery from the drone flights were processed (see below), we undertook an additional visual inspection on foot to confirm entrances identified from the thermal imagery and to identify any false-positives or -negatives.

Prior to undertaking the surveys, we flew each imager at various flight heights and speeds to determine optimum picture quality. For the survey, the Zenmuse was flown at 3 m/s at 10m above ground level (AGL).  This resulted in a swath width of 15.6m and a resolution of 1.4 pixels/cm.  The Vayu was flown at 5 m/s at 40m AGL resulting in a swath width of 39.8m and a resolution of 2 pixels/cm.  Transect spacing for each imager for the survey was determined by the swath width.  Transects spacing for the Zenmuse was 11m resulting in a transect overlap of 2.3m either side of the image.  Sixteen transects were required to cover the area taking two flights to complete.  Transect spacing for the Vayu was 22m resulting in a transect overlap of 8.5m.  Eight transects were required to cover the area which took one flight to complete.  For both the Zenmuse and the Vayu the imager was pointed 90 degrees to the horizontal during the surveys.  During the surveys proper the drone was not stopped over entrances or warrens for confirmation of detection.

We downloaded the footage from the thermal imagers to an external hard drive and reviewed the footage from this drive using VLC media player 3.0.8. We recorded observations in a custom-built Microsoft Excel (Microsoft Corporation 2018) workbook which utilised the drone’s tracklog to georeference observation locations.  This file was then exported as a KML file and viewed in Google Earth Pro (Google Earth Pro 2019) to aid in comparison between thermal imager and visual inspection detections.

Where transect imagery overlapped, double observations of warren entrances were removed from the worksheet before analysis.  If a warren complex was identified on one transect, and additional warren entrances were identified on the immediate next transect in the same location, then a determination was made on whether these entrances belonged to the same warren or constituted a new warren.  This ensured warren counts were not over-estimated. 

Warrens were classified by the amount of vegetation present that was likely to obscure entrances.  Warrens with no vegetation present were classified as “open”, warrens obscured by vegetation (e.g. entrances were beneath shrubs) were classified as “vegetated” and warrens that had entrances in the open and obscured by vegetation were classified as “mixed”.  These classifications also applied to the entrances associated with that warren for analysis (i.e. individual entrances in “mixed” warrens were not further classified into “open” or “vegetated” categories for analysis).

Statistical analysis

We used the lme4 (Bates, Maechler et al. 2015) and lmerTest (Kuznetsova, Brockhoff et al. 2017) packages in R (R Core Team 2019) to test for any difference in entrance count associated with imager. We used a mixed model with Poisson likelihood to account for the nested structure of imagers within warrens and the contrast of vegetation class between distinct warren sets.  Package emmeans (Lenth 2019) was used to inspect the mean entrance count under each vegetation and imager class.  Additionally, we plotted difference between estimates vs average of the estimates to check for any patterning in case agreement depended on magnitude of observation as suggested by Altman and Bland (1983).  To address any disagreement in terms of presence or absence of entrances detected, the three pairings of methods (visual vs Vayu, visual vs Zenmuse and Vayu vs Zenmuse) were examined by classifying entrance counts as equal to or greater than zero and forming two-way tables.

热成像技术（thermal imaging technology）是野生动物管理领域的新兴发展方向。目前野生动物科学领域的热成像应用研究大多局限于大型有蹄类动物（ungulates）以及地表栖息的哺乳动物，针对使用热成像仪探测穴居动物（fossorial animals）及其洞穴的相关研究则相对匮乏。诸如白光探照灯这类常规调查方法，往往难以识别洞穴（进而无法探测到其中栖息的动物），在植被遮挡洞穴的区域这一问题尤为突出。热成像仪可通过捕捉洞穴的辐射热量来探测其中栖息的动物，在植被覆盖区域的探测效果尤为显著。如今，小型化、高性价比的热成像设备愈发普及，结合无人机（drone）技术后，更为研究人员与土地管理者将该技术应用于科研与管理实践提供了可能。

本研究对比了挂载于无人机的消费级（售价低于20000澳元）与专业级（售价高于65000澳元）热成像仪对野外兔穴（warrens）及其入口的探测能力，并与人工目视检查结果进行对照。

在植被稀疏的区域，热成像与目视检查均可有效探测到活跃兔穴。植被覆盖情况对兔穴入口的探测成功率存在显著影响（P<0.001，α=0.05）。消费级热成像仪探测到的兔穴入口数量显著少于专业级热成像仪与目视检查（P=0.009，α=0.05）。被植被遮挡的活跃兔穴入口无法通过消费级热成像仪导出的影像准确识别，且在回放该类影像时会出现较多假阳性检测结果。

研究认为，可导出帧率（Hz）是影响成像质量与后续假阳性检测的关键因素，在选型热成像仪时应重点考量该参数，建议最低可导出帧率需达到30Hz。热成像仪可作为辅助工具，有效提升活跃兔穴入口的识别效率；其中专业级热成像仪探测到的兔穴与入口数量均多于消费级热成像仪与人工目视检查。

研究方法
研究使用了三款不同传感器尺寸与价位的非制冷微测辐射热计阵列（uncooled microbolometer arrays，详见表1）。第一部分实验采用Jenoptik VarioCAMⓇ HD（以下简称"Jenoptik"）专业热成像仪；第二部分实验则使用FLIR Zenmuse XT640与Sierra-Olympic VayuHD，分别简称为"Zenmuse"与"Vayu"。其中Zenmuse为与DJI Inspire 1无人机集成的一体式系统，而Jenoptik与Vayu均为重量更大的非集成式热成像仪，二者均需挂载于Ronin MX云台（https://www.dji.com/au/ronin-mx）以实现影像稳定。Jenoptik搭载于DJI S1000+无人机（https://www.dji.com/au/spreading-wings-s1000/spec），Vayu则搭载于DJI Matrice 600无人机（https://www.dji.com/au/matrice600/info#specs）。所有视频均以"白热"灰度模式采集与处理。

在高密度兔群区域，区分不同兔穴对应的入口存在一定难度。本研究规定：若两个兔穴入口间距小于5m，则视为同一兔穴的入口；若某入口与其他入口间距超过5m，则视为独立新兔穴的入口；间距超过5m的单个入口则归类为单入口兔穴。当兔穴存在至少一个有使用痕迹的入口时，即判定为活跃兔穴，其判定依据包括入口处无植被生长、存在新挖掘的泥土、新鲜粪便以及兔脚印等。若兔穴所有入口均被碎屑（落叶、树枝）、蛛网与硬壳状土壤覆盖，则视为非活跃兔穴。本研究未通过挖掘或诱捕等方式进一步验证兔穴的活跃状态。

所有热成像调查均于日出前的清晨开展，以最大化兔穴入口与周边环境的温度差。在热成像调查前，研究人员于日间徒步对所有样地开展兔穴（活跃与非活跃）目视检查，并记录所有已识别兔穴的GPS坐标。地面调查与空中调查相互独立，即热成像仪的航线规划先于目视检查完成。第一部分实验旨在验证热成像仪能否探测到活跃兔穴，研究人员操控搭载Jenoptik的无人机直接飞往已标记的兔穴位置。第二部分实验则对比专业级热成像仪（Vayu）与消费级热成像仪（Zenmuse）的探测效果：研究人员规划了平行飞行航线以实现调查区域的全覆盖，并模拟实际野外兔穴搜索的调查流程。第二部分实验同时开展了兔穴与入口的目视计数：研究人员在抵达样地后、无人机飞行前开展目视计数，沿间距约10m的平行线路徒步记录所有兔穴及其关联入口。待无人机飞行获取的影像完成处理后，研究人员再次徒步开展目视检查，以确认热成像影像中识别出的入口，并统计假阳性与假阴性检测结果。

正式调查前，研究人员在不同飞行高度与速度下开展试飞，以确定最优成像参数。本次调查中，Zenmuse的飞行速度为3m/s，飞行高度为离地10m（AGL），对应的扫描幅宽为15.6m，分辨率为1.4像素/厘米。Vayu的飞行速度为5m/s，飞行高度为离地40m（AGL），扫描幅宽为39.8m，分辨率为2像素/厘米。两款热成像仪的航线间距由扫描幅宽决定：Zenmuse的航线间距为11m，航线两侧重叠宽度为2.3m，需执行16条航线、两次飞行完成全部区域覆盖；Vayu的航线间距为22m，航线两侧重叠宽度为8.5m，仅需8条航线、一次飞行即可完成全部区域覆盖。两款热成像仪在调查过程中均保持与水平方向成90°的俯拍角度，且飞行过程中未在兔穴入口或兔穴上空悬停以确认检测结果。

研究人员将热成像仪采集的视频下载至外置硬盘，通过VLC media player 3.0.8回放并审阅视频内容，将观测结果记录于定制的Microsoft Excel（Microsoft Corporation 2018）工作簿中，该工作簿利用无人机的飞行轨迹日志实现观测位置的地理配准。随后将该文件导出为KML格式，在Google Earth Pro（Google Earth Pro 2019）中查看，以辅助对比热成像检测结果与目视检查结果。

对于航线重叠区域采集的影像，需剔除重复观测的兔穴入口数据后再开展分析。若某兔穴群在一条航线上被识别，且相邻航线的同一位置存在额外入口，则需判断这些入口属于同一兔穴还是新兔穴，以避免兔穴计数被高估。

研究依据遮挡兔穴入口的植被覆盖情况对兔穴进行分类：无植被遮挡的兔穴归类为"开阔型"，入口被植被（如灌木下方）遮挡的兔穴归类为"植被覆盖型"，同时存在开阔与被植被遮挡入口的兔穴归类为"混合型"。该分类标准同样适用于关联兔穴的入口分析（即"混合型"兔穴的单个入口无需进一步划分为"开阔"或"植被覆盖"类别）。

统计分析
本研究使用R（R Core Team 2019）中的lme4（Bates, Maechler et al. 2015）与lmerTest（Kuznetsova, Brockhoff et al. 2017）包，检验不同热成像仪的入口探测数量是否存在差异。研究采用泊松似然混合模型，以适配兔穴内热成像仪的嵌套结构以及不同兔穴组间的植被类别差异。使用emmeans包（Lenth 2019）分析不同植被类别与热成像仪类别下的平均入口探测数量。此外，参照Altman与Bland（1983）的建议，绘制估计值差值与估计值平均值的散点图，以检查观测一致性是否随观测量级变化。针对探测入口的存在与否的分歧，将三种方法配对（目视检查 vs Vayu、目视检查 vs Zenmuse、Vayu vs Zenmuse）的入口计数按是否≥0进行分类，并构建双向列联表开展分析。