Deep object detection for waterbird monitoring using aerial imagery Krish Kabrayz Alexander Xiongyz Wenbin Liyz Minxuan Luoz William Luz

2025-05-06 0 0 5.93MB 7 页 10玖币

侵权投诉

Deep object detection for waterbird monitoring

using aerial imagery

Krish Kabra∗,†,‡, Alexander Xiong†,‡, Wenbin Li†,‡, Minxuan Luo‡, William Lu‡,

Tianjiao Yu‡, Jiahui Yu‡, Dhananjay Singh‡, Raul Garcia‡, Maojie Tang‡,

Hank Arnold§, Anna Vallery§, Richard Gibbons¶, Arko Barman§

‡Rice University, Houston, TX 77005, USA

§Houston Audubon Society, Houston, TX 77079, USA

¶American Bird Conservancy, The Plains, VA 20198, USA

Abstract—Monitoring of colonial waterbird nesting islands is

essential to tracking waterbird population trends, which are

used for evaluating ecosystem health and informing conservation

management decisions. Recently, unmanned aerial vehicles, or

drones, have emerged as a viable technology to precisely monitor

waterbird colonies. However, manually counting waterbirds from

hundreds, or potentially thousands, of aerial images is both

difﬁcult and time-consuming. In this work, we present a deep

learning pipeline that can be used to precisely detect, count,

and monitor waterbirds using aerial imagery collected by a

commercial drone. By utilizing convolutional neural network-

based object detectors, we show that we can detect 16 classes of

waterbird species that are commonly found in colonial nesting

islands along the Texas coast. Our experiments using Faster

R-CNN and RetinaNet object detectors give mean interpolated

average precision scores of 67.9% and 63.1% respectively.

Index Terms—Object detection, Convolutional neural net-

works, Wildlife monitoring

I. INTRODUCTION

Colonial waterbird nesting islands can be found across

the globe, and each of North America’s coasts is home to

its own species of breeding colonial waterbirds. Colonial

waterbirds are important indicators of ecosystem health [1],

provide numerous ecosystem services, and are an important

part of a growing nature-based tourism sector of the economy

[2]. Therefore, continuing research and monitoring of these

species is critical to inform conservation decisions, encourage

management of habitats for the beneﬁt of colonial waterbirds,

and to continue to gauge the surrounding ecosystem health.

Monitoring of waterbirds at colonial nesting islands is a

widespread technique used to track population trends. There

are many colonial waterbird monitoring programs in the U.S.

including the Texas Colonial Waterbird Survey, which is one of

the longest running programs that monitors waterbirds across

the entire Texas coast annually since 1976 [3]. Censusing

waterbirds on islands is no small task. Traditional monitoring

studies of waterbirds have been conducted by traversing the

colony on foot, surveying via boat, or surveying aerially using

small, manned aircraft. Each of these methods has its own

set of challenges and consequences. Surveying waterbirds by

foot can disturb both the species of interest and the habitat

†Denotes equal contribution. ∗Corresponding author: kk80@rice.edu

occupied. Low vantage points of boat-based surveys can result

in the risk of missing nests, particularly on larger and higher

islands. Moreover, in certain conditions, accessing the islands

by boat can be tricky due to inclement weather conditions.

Manned aerial surveys is the preferred technique by state and

federal wildlife agencies, but these are expensive and require

the proper conditions. In fact, aircraft crashes and boating

accidents have been found to be the largest causes of mortality

and injury among biologists in the ﬁeld [4].

In recent years, unmanned aerial vehicles (UAVs), also

referred to as drones, have presented themselves as a useful

tool in wildlife management [5], [6], including waterbird

monitoring [7], [8]. Drones allow researchers to remain safely

on the ground while surveying areas of interest with both

less cost and greater ease than traditional aerial surveys. In

studies where this technology has been applied, the use of

drones was found to result in more precise count estimates

than traditional ground-based surveys [9]. Unfortunately, the

expertise and time required to manually localize and classify

species from hundreds, potentially thousands, of aerial images

represents a major bottleneck.

To alleviate this issue, we developed a object detection-

based deep learning pipeline that utilizes convolutional neural

networks (CNNs) [10]–[12] to precisely localize and classify

colonial waterbird species from UAV aerial imagery via super-

vised learning. We collect survey images from three colonial

nesting islands along the Texas coast, and train a CNN-based

object detection model to detect 16 classes of waterbirds, in-

cluding the 14 most common colonial waterbird species found

on these islands: Brown Pelican (Pelecanus occidentalis),

Laughing Gull (Leucophaeus atricilla), Royal Tern (Thalasseus

maximus), Sandwich Tern (Thalasseus sandvicensis), Great

Egret (Ardea alba), Cattle Egret (Bubulcus ibis), Snowy Egret

(Egretta thula), Reddish Egret (Egretta rufescens), American

White Ibis (Eudocimus albus), Great Blue Heron (Ardea

Herodias), Black-crowned Night Heron (Nycticorax nyctico-

rax), Tri-colored Heron (Egretta tricolor), Roseate Spoonbill

(Platalea ajaja), and Black Skimmer (Rynchops niger). We

present results using two of the most commonly implemented

CNN-based object detection models, Faster R-CNN [11] and

RetinaNet [12].

arXiv:2210.04868v2 [cs.CV] 13 Oct 2022

A. Contributions

The key contributions of this work are as follows:

•We develop a deep learning pipeline to detect water-

birds from UAV aerial imagery for precise waterbird

monitoring. Our pipeline is general and can be applied

to other applications requiring object detection from

high-resolution aerial imagery, including other wildlife

monitoring applications. Our code is available at: https:

//github.com/RiceD2KLab/Audubon F21

•We apply our method to detect 16 classes of waterbirds

from UAV aerial imagery collected from nesting islands

surveyed along the Texas coast. To the best of the authors’

knowledge, this is one of the largest number of species

detected by a single model for UAV-based waterbird

monitoring research.

•We present experimental results utilizing Faster R-CNN

and RetinaNet object detectors. We show that we can

accurately detect 3 of the most prevalent waterbird classes

in our dataset (>70% of total waterbirds): Mixed Tern

Adults, Laughing Gull Adults, and Brown Pelican Adult,

with an interpolated average precision (APIoU=0.5) score

of over 90% for Faster R-CNN and 85% for RetinaNet.

Across all waterbird classes, we achieve a mean inter-

polated average precision (mAPIoU=0.5) scores of 67.9%

and 63.1% for Faster R-CNN and RetinaNet respectively.

II. RELATED WORK

In recent years, object detection, the task of localizing one

or more objects in an image with corresponding classiﬁcations,

has seen immense advancements largely due to the rapid

development of deep learning [13]. State-of-the-art object

detection architectures utilizing convolutional neural networks

(CNN) as a ‘backbone’ have particularly achieved much

success due to a CNN’s ability to learn hierarchical image

features [14], [15]. For this work we utilize two popular CNN-

based object detectors: Faster R-CNN [11] and RetinaNet [12].

Nevertheless, the proposed method is general, and is easily

extensible to other object detectors.

Consequent to the success of CNN-based object detection

and wide availability of open-source code, several works have

utilized these methods for wildlife monitoring with unmanned

aerial vehicle (UAV) imagery. Andrew et al. [16] use a R-

CNN [10] to detect Holstein Friesian cattle from UAV imagery,

proposing both a standard still-image acquisition pipeline and

an extended video monitoring pipeline. Kellenberger et al.

[17] use a custom one-stage detector with an AlexNet [14]

backbone to detect large animals from UAV images captured

over the Kuzikus wildlife reserve park in Namibia. Gray et al.

[18] use a Mask R-CNN [19] to detect and segment humpback

whales, minke whales, and blue whales from UAV imagery

collected off the coast of California and along the Western

Antarctic Peninsula.

More related to this work, researchers have also utilized

CNN-based object detectors for bird monitoring. As compared

to the aforementioned works, which focused on detecting

Fig. 1. Dataset distribution of waterbird classes. 24 classes of various bird

species at different ages and conﬁgurations. In this work, we select 15 unique

classes to be detected by the object detector, and combine the remaining

classes into the “Other” waterbird class (highlighted by italicized, blue text).

relatively large and distinct mammals, bird detection from

UAV imagery is generally regarded as a more challenging

detection task due to the unique characteristics of birds. In

particular, visual differences between bird species may be

minor, making it difﬁcult to distinguish between them. This

difﬁculty is heightened for partially occluded birds, such as

birds with necks tucked under their wings, as key visual

features used to make distinctions are hidden. Borowicz et

al. [20] use DetectNet [21] to count Ad´

elie penguins in the

Danger Islands off the northern tip of the Antarctic Peninsula.

Hong et al. [22] survey various off-the-shelf CNN-based object

detectors to detect birds from UAV imagery collected of

both wild and decoy birds in various environments across

South Korea. Hayes et al. [23] use a RetinaNet to detect

seabirds, speciﬁcally Black-browed Albatrosses and Southern

Rockhopper Penguins, from UAV imagery collected of the

Falkland (Malvinas) Islands.

This work expands on the existing literature of deep

learning-based object detection for bird monitoring by sig-

niﬁcantly increasing the number of bird species detected by

a single object detector. The aforementioned works focus on

identifying 2 or fewer bird classes that are often speciﬁc to the

surveyed islands. However, this work shows that CNN-based

object detectors are capable of detecting several bird species,

even when visually similar or imaged in challenging viewing

conditions such as dense ﬂocks or obscuring foliage.

III. DATASET

Aerial imagery from three colonial waterbird nesting is-

lands, Chester Island, Little Bay North Island, and North Deer

文档加载中……请稍候！
如果长时间未打开，您也可以点击刷新试试。

下载文档到电脑，查找使用更方便

10 玖币 0人已下载

立即下载

摘要：

DeepobjectdetectionforwaterbirdmonitoringusingaerialimageryKrishKabra,y,z,AlexanderXiongy,z,WenbinLiy,z,MinxuanLuoz,WilliamLuz,TianjiaoYuz,JiahuiYuz,DhananjaySinghz,RaulGarciaz,MaojieTangz,HankArnoldx,AnnaValleryx,RichardGibbons{,ArkoBarmanxzRiceUniversity,Houston,TX77005,USAxHoustonAudubonSociety,...

展开>> 收起<<

Deep object detection for waterbird monitoring using aerial imagery Krish Kabrayz Alexander Xiongyz Wenbin Liyz Minxuan Luoz William Luz.pdf

共7页,预览2页

还剩页未读，继续阅读

声明：本站为文档C2C交易模式，即用户上传的文档直接被用户下载，本站只是中间服务平台，本站所有文档下载所得的收益归上传人(含作者)所有。玖贝云文库仅提供信息存储空间，仅对用户上传内容的表现方式做保护处理，对上载内容本身不做任何修改或编辑。若文档所含内容侵犯了您的版权或隐私，请立即通知玖贝云文库，我们立即给予删除！

Deep object detection for waterbird monitoring using aerial imagery Krish Kabrayz Alexander Xiongyz Wenbin Liyz Minxuan Luoz William Luz

相关推荐

开通VIP享超值会员特权

作者详情

相关内容

热门标签

举报选择: