Search USGS⌕ Search
Home/Archive/USGS

USGS · 70246285

Challenges and solutions for automated avian recognition in aerial imagery

Zhonqgi Miao·Stella X Yu·Kyle Lawrence Landolt·Mark D. Koneff·Timothy White·Luke J. Fara·Enrika Hlavacek·Bradley A. Pickens·Travis J. Harrison·Wayne M. Getz
View original source ↗Export metadata ↓

Abstract

Remote aerial sensing provides a non-invasive, large geographical-scale technology for avian monitoring, but the manual processing of images limits its development and applications. Artificial Intelligence (AI) methods can be used to mitigate this manual image processing requirement. The implementation of AI methods, however, has several challenges: (1) imbalanced (i.e., long-tailed) data distribution, (2) annotation uncertainty in categorization, and (3) dataset discrepancies across different study sites. Here we use aerial imagery data of waterbirds around Cape Cod and Lake Michigan in the United States to examine how these challenges limit avian recognition performance. We review existing solutions and demonstrate as use cases how methods like Label Distribution Aware Marginal Loss with Deferred Re-Weighting, hierarchical classification, and FixMatch address the three challenges. We also present a new approach to tackle the annotation uncertainty challenge using a Soft-fine Pseudo-Label methodology. Finally, we aim with this paper to increase awareness in the ecological remote sensing community of these challenges and bridge the gap between ecological applications and state-of-the-art computer science, thereby opening new doors to future research.

Explore related subjects

Remote Sensing in Ecology and ConservationArticle

Keep this discovery

Explore connections, maps & timelines

BibTeXRIS

Zhonqgi Miao, Stella X Yu, Kyle Lawrence Landolt, Mark D. Koneff, Timothy White, Luke J. Fara, Enrika Hlavacek, Bradley A. Pickens, Travis J. Harrison, Wayne M. Getz. 2023-04-26. Challenges and solutions for automated avian recognition in aerial imagery. https://doi.org/10.1002/rse2.318

Cite the original work for its findings. Save a collection to share your selection of sources.

KEEP EXPLORING

Related USGS reports

USGS2025

Object detection-assisted workflow facilitates cryptic snake monitoring

Camera traps are an important tool used to study rare and cryptic animals, including snakes. Time-lapse photography can be particularly useful for studying snakes that often fail to trigger a camera's infrared motion sensor due to their ectothermic nature. However, the large datasets produced by time-lapse photography require labor-intensive classification, limiting their use in large-scale studies. While many artificial intelligence-based object detection models are effective at identifying mammals in images, their ability to detect snakes is unproven. Here, we used camera data to evaluate the efficacy of an object detection model to rapidly and accurately detect snakes. We classified images manually to the species level and compared this with a hybrid review workflow where the model removed blank images followed by a manual review. Using a ≥0.05 model confidence threshold, our hybrid review workflow correctly identified 94.5% of blank images, completed image classification 6× faster, and detected large (>66 cm) snakes as well as manual review. Conversely, the hybrid review method often failed to detect all instances of a snake in a string of images and detected fewer small (<66 cm) snakes than manual review. However, most relevant ecological information requires only a single detection in a sequence of images, and study design changes could likely improve the detection of smaller snakes. Our findings suggest that an object detection-assisted hybrid workflow can greatly reduce time spent manually classifying data-heavy time-lapse snake studies and facilitate ecological monitoring for large snakes.

Florida↗
USGS2023

High-resolution thermal imagery reveals how interactions between crown structure and genetics shape plant temperature

Understanding interactions between environmental stress and genetic variation is crucial to predict the adaptive capacity of species to climate change. Leaf temperature is both a driver and a responsive indicator of plant physiological response to thermal stress, and methods to monitor it are needed. Foliar temperatures vary across leaf to canopy scales and are influenced by genetic factors, challenging efforts to map and model this critical variable. Thermal imagery collected using unoccupied aerial systems (UAS) offers an innovative way to measure thermal variation in plants across landscapes at leaf-level resolutions. We used a UAS equipped with a thermal camera to assess temperature variation among genetically distinct populations of big sagebrush ( Artemisia tridentata ), a keystone plant species that is the focus of intensive restoration efforts throughout much of western North America. We completed flights across a growing season in a sagebrush common garden to map leaf temperature relative to subspecies and cytotype, physiological phenotypes of plants, and summer heat stress. Our objectives were to (1) determine whether leaf-level stomatal conductance corresponds with changes in crown temperature; (2) quantify genetic (i.e., subspecies and cytotype) contributions to variation in leaf and crown temperatures; and (3) identify how crown structure, solar radiation, and subspecies-cytotype relate to leaf-level temperature. When considered across the whole season, stomatal conductance was negatively, non-linearly correlated with crown-level temperature derived from UAS. Subspecies identity best explained crown-level temperature with no difference observed between cytotypes. However, structural phenotypes and microclimate best explained leaf-level temperature. These results show how fine-scale thermal mapping can decouple the contribution of genetic, phenotypic, and microclimate factors on leaf temperature dynamics. As climate-change-induced heat stress becomes prevalent, thermal UAS represents a promising way to track plant phenotypes that emerge from gene-by-environment interactions.

Idaho↗
USGS2023

Both Landsat- and LiDAR-derived measures predict forest bee response to large-scale wildfire

Large-scale disturbances such as wildfire can have profound impacts on the composition, structure, and functioning of ecosystems. Bees are critical pollinators in natural settings and often respond positively to wildfires, particularly in forests where wildfire leads to more open conditions and increased floral resources. The use of Light Detection and Ranging (LiDAR) provides opportunities for quantifying habitat features across large spatial scales and is increasingly available to scientists and land managers for post-fire habitat assessment. We evaluated the extent to which LiDAR-derived forest structure measurements can predict forest bee communities after a large, mixed-severity fire. We hypothesized that LiDAR measurements linked to post-fire forest structure would improve our ability to predict bee abundance and species richness when compared to satellite-based maps of burn severity. To test this hypothesis, we sampled wild bee communities within the Douglas Fire Complex in southwestern Oregon, USA. We then used LiDAR and Landsat data to quantify forest structure and burn severity, respectively, across bee sampling locations. We found that the LiDAR forest structure model was the best predictor of abundance, whereas the Landsat burn severity model had better predictive ability for species richness. Furthermore, the Landsat burn severity model was better at predicting the presence and species richness of bumble bees (Bombus spp.), an ecologically distinct and economically important group within the Pacific Northwest. We posit that the divergent responses of the two modeling approaches are due to distinct responses by bee taxa to variation in forest structure as mediated by wildfire, with bumble bees in particular depending on closed-canopy forest for some portions of their life cycle. Our study demonstrates that LiDAR data can provide information regarding the drivers of bee abundance in post-wildfire conifer forest, and that both remote sensing approaches are useful for predicting components of wild bee diversity after large-scale wildfire.

Oregon↗