Oblique Aerial Surveys

KEY DETAILS

Principal Investigator
Dr. Matthew Rogan
Date
16 April 2024
Version
0.1.0
Programme
Rangelands Biodiversity Project
Study Site
Lewa-Lolldaiga-Borana-Ngare Ndare (LLBN)
Key partners
Ndoto Drones
Contact email
mrogan@naturalstate.org

1. PREAMBLE

Natural State’s objectives and activities are governed by a set of accepted Design Documents (DDs). These documents describe the context and purpose of all Natural State projects. Each DD documents key project details, the objective and background of the project, features of the study area, and the general methodological framework. Specific methodological details may be found in the project Standard Operating Procedures (SOP) which is available in the Related Documents section below.

1.1 DD PURPOSE

To provide a clear understanding of the purpose of each Natural State project and its contribution to Natural State’s mission of facilitating nature restoration at scale by using the latest technology and methods to revolutionise impact monitoring for carbon, biodiversity and human well-being.

1.2 DD SCOPE

This document details how this project fits within Natural State’s Impact Monitoring strategy and the principal team members overseeing the project. It explains why the project was conceived and how it will be implemented. It further directs readers to where they can find additional information relevant to the project.

2. GLOSSARY

AGL
Above Ground Level. The height, in meters, of an aerial vehicle measured at any given time relative to the nadir.
AOI
Area of interest. A specific area within the landscape for which the aim is to collect a representative sample of biodiversity.
BACI
Before-After Control-Impact survey design.
Camera angle
The angle of the camera relative to the horizontal flightpath of the UAV such that 90° is straight down (i.e. nadir).
Exclosure
A defined area fenced intended to prevent some species of wild animals from accessing the area to reduce browsing and prevent unsafe human-wildlife encounters or facilitate tree growth and recruitment. Exclosures have a variety of fencing types that restrict the movement of either elephants and giraffe or all medium-large herbivores. In actuality, excosures reduce access but do not prevent access to excluded animals.
Extension
An exclosure or portion of exclosure erected in 2022 or later.
Image overlap
The percent of two sequential images that is captured in both.
LLBN
The Lewa-Lolldaiga-Borana-Ngare Ndare landscape (See study site description for details).
Nadir
The point on the ground closest (i.e., directly below) to the aerial vehicle.
Oblique aerial imagery
Still photographs of the landscape taken from an aerial vehicle at an angle such that neither the horizon nor the nadir point below the aerial vehicle are in the field of view.
Open grassland
Habitat patches with few or no trees, or a small clump of trees highly localized within the patch or immediate vicinity.
Open Woodland
Habitat patches that are covered primarily by grass but trees or bushes are spaced somewhat regularly throughout most of the patch or immediate vicinity.
Project
A concerted, data-driven effort to robustly measure variation in Biodiversity, Carbon, or Human-wellbeing in response to one or more sources of heterogeneity in a designated landscape.
Sampling design
The set of field methods employed in a survey and the manner of their use.
Sampling protocol
Explicit survey methodology that describes the design, effort, duration, configuration, and operation of a survey.
Site, Sampling
A distinct, discrete spatial unit defined in at least two dimensions where sampling occurs.
Study area
A defined geographic region of interest within which one or more surveys investigate ecological patterns at one or more sites.
Survey
A set of simultaneous deployments or related deployment groups of remote sensors over a defined period of time at a coordinated set of stations for the purposes of collecting data on the environment and its biological communities as part of a NS project.
Survey design
The theoretical and practical methods for choosing the spatiotemporal distribution of sampling units in a survey.
Transect
A one-dimensional, discrete spatial unit where sampling occurs.
Treatment
Spatial management effects with ‘control’ indicating the absence of the management effect.
UAV
Unpiloted Aerial Vehicle
Vegetation structure
Variation in height and woodiness of vegetation that extends above the grass.

3. PROJECT OVERVIEW

3.1 PROJECT AIMS

The Oblique Aerial Imagery project aims to:

  1. Collect photographic observations of at least six and preferably 12 species of large mammal for the purpose of training a machine learning classifer to predict species classes.
  2. Conduct a baseline survey of large mammal populations in areas of recent or planned expansions of exclosures and adjacent control areas.
  3. Provide real world data to inform future survey design based on observation rates, sampling variance, and other factors.
  4. Investigate the effect of tree cover on species observation rates.

3.2 PROJECT BACKGROUND

Natural State is focused on catalysing large-scale restoration globally by revolutionizing impact monitoring, developing new financial mechanisms and supporting local leaders. As part of this, NATURAL STATE aims to develop and optimise impact monitoring systems to reduce the cost and provide independently verifiable results on interventions. Large scale restoration requires large scale monitoring, for which aerial census is widely used. An area of active development is the use of imagery from aerial platforms, combined with machine learning classifiers to provide fast and more cost-effective records of observations. With appropriate statistical inference, this field of research can provide robust and defensible estimates of abundance of large mammals, and their change through time. Aerial imagery is more widely used, especially as new platforms such as unpiloted aerial vehicles (UAVs) become more effective and available. However, several key questions exist about how best to optimize camera and survey configuration, and regarding the effect of vegetaion on animal detection. Questions also remain about the performance of machine learning classifiers for aerial imagery and the most robust statistical frameworks for inferring abundance from aerial observations.

Historically, aerial surveys have relied on human observers “spotting” animals and reporting observations to a data recorder. However, human observations are labor intensive, expensive, imprecise, and error prone. Direct human observations are also incompatible with UAVs. In recent years, computer vision algorithms have been used for object recognition of animals in aerial imagery (e.g., Eikelboom et al. 2019). Such machine learning tools offer much greater efficiency, reproducibility, and scalability than traditional human-observer approaches. Nevertheless, algorithms with high recall and high precision for out-of-sample mammal detection and identification are not currently available. Most recent developments focus on rapid annotation and model training (e.g., Kellenberger et al. 2018; Zheng et al. 2021) rather than generalizable pre-trained algorithms. Additionally, obscured visibility due to tree cover or uneven terrain can further degrade classifier performance.

NATURAL STATE is planning to tackle those challenges for aerial surveys conducted using UAVs. A UAV will fly low-altitude (~50 m AGL), temporally and spatially replicated transects. While the primary aim of the surveys is to generate a training dataset, they will nevertheless be designed to provide representative samples of the AOIs. The surveys will yield imagery with a resolution of no more than 2 cm at the image centroid.

3.3 STUDY AREA

The Lewa-Lolldaiga-Borana-Ngare Ndare (LLBN) study site is a savanna rangeland landscape in the central highlands of Kenya that extends north from the lower slopes of Mt. Kenya, straddling the boundary between Meru and Laikipia counties. The four main properties are located at 0.11 = 0.34° latitude and 37.07-37.53° longitude. The study site-centered ecosystem (SCE) exhibits three main “arms” with on extending south around the western edge of Mt. Kenya, one extending east along the southern side of the B9 highway to the northeastern edge of Meru County, and the longest and largest arm extending 130 km northwest to Maralal with some isolated islands of similar ecosystem to the north and west. The entire SCE is located between -0.3° and 2.3° latitude, and between 36.3° and 38.1° longitude.

Rainfall is highly variable, but is typically between 400 and 600 mm annually. During droughts, total annual rainfall can drop below 200 mm. Rainfall generally follows a north-south gradient with more rainfall at higher elevations in the south. The landscape sits at 1400-2370 m with higher elevations in Ngare-Ndare forest and south-central Lolldaiga and the lowest elevations in Lewa and northeast Borana. Soil types are highly variable throughout the landscape but consist primarily of Luvisols (Haplic and Vertic) and Dystric Regosols. Vegetation communities are predominantly Acacia-Commiphora bushlands and thickets with montane forests at higher elevations (Dinerstein et al. 2017).

Each of the four core properties are subject to divergent management practices. Lewa Wildlife Conservancy is mostly owned by The Nature Conservancy but has a number of small, privately owned enclaves. The reserve is almost exclusively used for photographic tourism but does have some livestock grazing, especially along the community road and in a section nicknamed ‘Bosnia’ where the community have grazing rights. Over the last 30 years, the conservancy has increasingly adopted elephant and large-mammal exclosures as a major management practice aimed at increasing tree cover on the conservancy.

Lolldaiga ranch is principally a livestock ranch but is also used by free ranging wildlife. Historically, the ranch was intensively grazed but recently new management has implemented a regenerative grazing regime based on intensive grazing by a few large, fast-moving herds. The grazing regime is intended to stimulate grass productivity and prevent selective grazing by livestock herds.

Borana Conservancy is a dual use wildlife and livestock ranch that also includes privately owned enclaves. In addition to photographic tourism, the conservancy produces cattle and has a partnership with a local community that extends limiting grazing rights to community herders. Borana has a few large-mammal exclosures, especially along the river that runs between Lewa and Borana.

Ngare-Ndare forest is a state forest reserve but is managed by the Ngare-Ndare Forest Trust on behalf of six villages. The Trust administers sustainable use of the forest by the six communities, predominantly for firewood collection and grazing. The forest also has small-scale tourism operations. The forest is predominantly populated with wild olives and junipers.

North of Lewa, Borana, and Lolldaiga are several community conservancies that practice predominatly pastoral lifestyles as well as the Mukogodo Forest, which is a forest conservancy managed on behalf of four neighboring communities. On the western boundary of Lolldaiga Ranch is Kupona, a small experiemental plot for testing restoration interventions.

Initial surveys for the Oblique Aerial Surveys will focus on four Areas of Interest (AOI): The Matunda Extension (treatment) and an adjacent control zone immediately northwest, the Kona Safi Extension (treatment) and an adjacent control zone immediately north, the Luai ya Charlie exclosure (treatment) and an adjacent control zone immediately east, and the nothern Lolldaiga plains.

3.4 PROJECT TIMELINE

The Matunda AOI will be surveyed for a minimum of three days in late April. The Kona Safi, Luai ya Charlie, and Lolldaiga AOIs will be surveyed for a minimum of three days each in May and June. Additional surveys will be conducted July - October.

4. SURVEY DESIGN

4.1 SPATIAL DESIGN

Each AOI will be surveyed via a set of parallel transects. Transects will be set at a width of 250m and be aligned with the primary axis of each AOI. Adjacent AOIs will be sampled using a single set of transects if they are small enough for the drone to fly at least two transects (i.e., out and back) on a single set of batteries.

4.2 TEMPORAL DESIGN

Each transect will be flown three times on three different days within a one-month period and preferably at three different times of day.

5. SAMPLING DESIGN

Oblique aerial surveys will be conducted using a DJI Mavic 3 Enterprise drone. The UAV will be programmed to take photographs at consistent intervals along the transect with 0% horizontal overlap and 10% vertical overlap between photos. The UAV will fly at a height of 40 m using a 20 MP or greater camera with a camera angle of 45°.

6. ANALYTICAL FRAMEWORK

Results will be analysed using raw counts and N-mixture models.

7. EXPECTED OUTPUTS

  1. Training datasets consisting of approximately 1000 photos for the following species: elephants, giraffe, rhinos (both species collectively), buffalo, eland, grevy’s zebra, common zebra, oryx, impala, grant’s gazelle, warthog.
  2. Representative baseline surveys of the Matunda and Kona Safi extensions and their respective control areas.
  3. Transect counts from northern Lolldaiga for direct comparison to aerial survey estimates acquired using fixed-wing aircraft.
  4. A comparison of object detection in the wooded habitat of Luai ya Charlie compared to the open grassland of the adjacent control area.
  5. Real-world date on capture rates in open grassland and open woodland habitat as well as estimates of sampling variance among replicates and among transects withing a given AOI.

8.1 STANDARD OPERATING PROCEDURE

None currently available.

8.2 OUTPUTS

None currently available.

8.3 DATA ELEMENTS

Survey Design

Data Collection

None currently available.

Data Layers

None currently available.

Dashboard

None currently available.

8.4 ADMINISTRATIVE DOCUMENTS

None currently available.

9. REVISION AND VERSION HISTORY AND DESCRIPTION

No history available.

10. SIGNATURES OF CONFIRMATION

Principal Investigator: ______________             Date: ___________

Director of Impact Insights: ____________             Date: ___________

11. BIBLIOGRAPHY

  1. Brack et al., 2022. Optimally designing drone-based surveys for wildlife abundance estimation with N-mixture models. Methods in Ecology and Evolution, 14, pp. 898-910.
  2. Delplanque et al. 2021. Multispecies detection and identification of African mammals in aerial imagery using convolutional neural networks. Remote Sensing in Ecology and Conservation, 8(2), 166-179.
  3. Eikelboom, J.A., et al. 2019. Improving the precision and accuracy of animal population estimates with aerial image object detection. Methods in Ecology and Evolution, 10(11), pp.1875-1887.
  4. Kellenberger, B., Marcos, D. and Tuia, D., 2018. Detecting mammals in UAV images: Best practices to address a substantially imbalanced dataset with deep learning. Remote sensing of environment, 216, pp.139-153.
  5. Kellenberger, B., Tuia, D., and Morris, D. 2020. AIDE: Accelerating image-based ecological surveys with interactive machine learning. Methods in Ecology and Evolution, 11, pp. 1716-1727.
  6. Lamprey et al. 2020. Comparing an automated high-definition oblique camera system to rearseat-observers in a wildlife survey in Tsavo, Kenya: Taking multi-species aerial counts to the next level. Biological Conservation, 241, 108243.
  7. Sundaram, N., and Meena, S.D. 2023. Integrated animal monitoring system with animal detection and classification capabilities: a review on image modality, techniques, applications, and challenges. Artificial Intelligence Review, 56, pp. 1-51.
  8. Zheng, X., Kellengerger, B., Gong, R., Hajnsek, I., and Tuia, D. 2021. Self-supervised pretraining and controlled augmentation improve rare wildlife recognition in UAV Images.

12. APPENDICES

None currently available