Engineering Papers⌕ Search

SEARCH · Engineering Papers

Results for “nighttime detection”

Search indexed NASA NTRS and DOE OSTI research on propulsion, heat transfer, battery materials and energy systems. Follow report and document links to the original sources.

Quote a phrase for an exact phrase match. Source license links do not imply unrestricted reuse.

16 records
All sourcesNASA NTRSDOE OSTI
DOE OSTI2022

Improving Single-Stage Object Detectors for Nighttime Pedestrian Detection

We report Improving the reliability of nighttime pedestrian detection is a crucial challenge towards the design of robust autonomous systems. Not surprisingly, most pedestrian fatalities occur in low-illumination settings, thus emphasizing the need for new algorithmic advances. This work presents a novel pedestrian detection approach that makes a number of crucial modifications to the state-of-the-art YOLOV5-PANet architecture, in order to improve the reliability of features extracted from nighttime images. More specifically, the proposed architecture systematically incorporates powerful shuffle attention mechanisms and a transformer module to improve the feature learning pipeline. Instead of advocating the use of other sensing modalities that are better suited for nighttime detection, our approach relies only on conventional RGB cameras and is hence broadly applicable. Our empirical studies with nighttime pedestrian detection benchmarks show that with only minimal increase in model complexity, our approach provides significant improvements in detection efficacy over existing solutions. Finally, we explore the impact of post-hoc network pruning on the speed-accuracy trade-off of our approach and demonstrate that it is well suited for reduced memory/compute requirements.

97 MATHEMATICS AND COMPUTING↗
DOE OSTI2023

Driving in the dark: Deciphering nighttime driver detection of free-ranging roadside wildlife

Wildlife-vehicle collisions are dangerous for motorists; however, few studies have addressed driver detection of roadside animals, and none have evaluated detection of free-ranging wildlife. Here, we used 24 volunteer drivers, infrared videography, a 75-km route, and free-ranging wildlife to quantify factors influencing (1) probability of wildlife detection, (2) detection distance, and (3) probability of dangerous encounters (i.e., detection distance < distance required for braking) for multiple species at night in South Carolina, USA. Detection probability of white-tailed deer (Odocoileus virginianus) was impacted by multiple driver, animal, and roadside factors. Deer detection distances increased by 20.99 m when drivers used high-beam headlights and 23.36 m when deer were moving but decreased by 0.71 m for every minute into a drive. Every encounter with wild pigs (Sus scrofa) and most encounters with small mammals were considered dangerous. Our findings suggest most drivers cannot safely detect deer, wild pigs, and small mammals at night.

54 ENVIRONMENTAL SCIENCES↗
DOE OSTI2024

Gaining Real-Time Water Leak Detection

Devens Reserve Forces Training Area is a United States Army Reserve (USAR) Installation that struggles with severe water leaks, often causing significant damage to the facility and requiring major renovation. Traditional water use is highly dependent on occupancy, so it can be difficult to benchmark a facility’s water use. It can be exceptionally difficult when occupancy is transient and/or varies. Pacific Northwest National Laboratory (PNNL) collaborated with Devens to implement real-time monitoring of their water consumption by utilizing the smart meter data from their existing 23 water meters. PNNL created a simple algorithm to calculate hourly water consumption and trigger an alert to be instantly emailed to Devens’ personnel when there appears to be a water leak in any building with a smart water meter. Here, this approach is expected to save hundreds of thousands of dollars in unnecessary water consumption costs and damages from leaks and was implemented with little-to-no costs or service disruptions. Next steps for this project include slow leak detection through nighttime monitoring and to extrapolate this water leak approach to the remainder 360 water meters on USAR’s Enterprise Building Control System so USAR sites across the country can be instantly notified of potential water leaks.

32 ENERGY CONSERVATION, CONSUMPTION, AND UTILIZATI↗
DOE OSTI2023

Automatic detection of ship-induced cloud features in satellite imagery

Ships crossing the ocean are known to produce long, curvilinear features called ship tracks visible in satellite imagery via the Twomey effect; however, there has been little exploitation of satellite imagery for broad atmospheric studies or global monitoring of ship emissions due to the difficulty of automated ship track detection. Prior studies are either proof-of-concept, qualitatively assessed, or restricted to a certain time of day. We propose a statistical method for the automated identification of ship tracks and demonstrate using GOES-West ABI data. We first present a human-assisted segmentation method, which we use to generate a ground truth data set of 529 annotated ship tracks in GOES-West ABI products. We then describe a two-stage automated approach comprising a detection stage to generate ship track proposals and a classification stage to reduce false positives. For detection, we present a novel pipeline based around a z-score filtering technique, and for classification, we demonstrate several classifiers from literature. In a final experiment, we quantitatively tune the detection parameters and train the classifier using the ground truth dataset, then test on a sequestered set of images; the detect-then-classify system had an overall Pd of 0.68 and 0.80 for daytime and nighttime data, respectively, and the classifier reduced false positive detections by 67% and 75%.

47 OTHER INSTRUMENTATION↗
DOE OSTI2025

A Multi-Sensor Approach for Measuring Bird and Bat Collisions with Offshore Wind Turbines (Final Technical Report)

Collision of birds and bats with wind turbines is a conservation concern for both land-based and offshore wind projects. The fatality rates of birds and bats at land-based turbines are well documented. The measurement strategies on land focus on finding carcasses following collision, estimating the number of carcasses missed through searcher efficiency, carcass persistence trials and carcass fall distributions, and modeling statistically robust fatality rates. Few technologies have been developed to monitor offshore bird and bat collisions, and many that have been developed focused on detecting collisions with large birds. The few studies that have attempted to document collisions at offshore turbines do not account for smaller bodied animals or for collisions that might be missed, which prevents the calculation of statistically robust fatality rates. The overall goal of this report, A Multi-Sensor Approach for Measuring Bird and Bat Collisions with Offshore Wind Turbines (Project), was to develop an effective multi-sensor system for quantifying bird and bat collision rates, specifically for offshore wind facilities. The Project goal and resulting automated collision detection system was achieved through two major technological advancements: 1) refining The Netherlands Organisation for Applied Scientific Research’s (TNO’s) existing WT-Bird® vibration sensing system, that had successfully detected large bird collisions during daytime, to allow for improved detection of smaller birds and bats during both daytime and nighttime hours and 2) improving image processing systems and developing and integrating machine learning algorithms to automatically detect and classify small and large bird and bat collisions with offshore turbines. This final technical report (FTR) summarizes Methods , Results , Conclusions , and Lessons Learned during each of the five Tasks identified for this research and development effort. This FTR includes summaries of the following: Task 1. Initial Engineering Tests to Improve WT-Bird® Task 2. Installation of WT‐Bird® on a Utility-scale Turbine at the National Wind Technology Center – National Renewable Energy Laboratory Task 3. Field Tests and Refinement of the Object Detection System Task 4. Validation of WT-Bird® on a Land-based Turbine Task 5. Preparation for the Implementation of WT-Bird® on an Offshore Turbine. This research and development effort documented successful improvement of the WT Bird® collision detection system to detect small birds and bats, and WT-Bird® is the first collision detection system to validate results compared to land-based post-construction monitoring. The collision trials provide estimates of missed targets that can be used to estimate fatality rates, a significant improvement relative to other offshore collision monitoring systems. Advances were made in developing an edge-processing solution to reduce data storage requirements, which is important if the system is deployed for long periods of time at offshore turbines. The improved WT-Bird® system also provides an important option for wind operators on land or offshore who need to document specific details about when collisions occur, particularly efforts to further research on bat impact minimization, or when standard fatality searches are impractical (e.g. offshore) or inadequate (e.g. challenging locations on land).

17 WIND ENERGY↗
DOE OSTI2022

Comparison of scattering ratio profiles retrieved from ALADIN/Aeolus and CALIOP/CALIPSO observations and preliminary estimates of cloud fraction profiles

The space-borne active sounders have been contributing invaluable vertically resolved information of atmospheric optical properties since the launch of Cloud-Aerosol Lidar and Infrared Pathfinder Satellite Observation (CALIPSO) in 2006. To build long-term records from space-borne lidars useful for climate studies, one has to understand the differences between successive space lidars operating at different wavelengths, flying on different orbits, and using different viewing geometries, receiving paths, and detectors. In this article, we compare the results of Atmospheric Laser Doppler INstrument (ALADIN) and Cloud-Aerosol Lidar with Orthogonal Polarization (CALIOP) lidars for the period from 28 June to 31 December 2019. First, we build a dataset of ALADIN–CALIOP collocated profiles (Δdist<1°; Δtime<6 h). Then we convert ALADIN's 355 nm particulate backscatter and extinction profiles into the scattering ratio vertical profiles SR(z) at 532 nm using molecular density profiles from Goddard Earth Observing System Data Assimilation System, version 5 (GEOS-5 DAS). And finally, we build the CALIOP and ALADIN globally gridded cloud fraction profiles CF(z) by applying the same cloud detection threshold to the SR(z) profiles of both lidars at the same spatial resolution. Before comparing the SR(z) and CF(z) profiles retrieved from the two analyzed lidar missions, we performed a numerical experiment to estimate the best achievable cloud detection agreement CDA norm (z) considering the differences between the instruments. We define CDA norm (z) in each latitude–altitude bin as the occurrence frequency of cloud layers detected by both lidars, divided by a cloud fraction value for the same latitude–altitude bin. We simulated the SR(z) and CF(z) profiles that would be observed by these two lidars if they were flying over the same atmosphere predicted by a global model. By analyzing these simulations, we show that the theoretical limit for CDA$_{norm}^{theor}$(z) for a combination of ALADIN and CALIOP instruments is equal to 0.81±0.07 at all altitudes. In other words, 19 % of the clouds cannot be detected simultaneously by two instruments due to said differences. The analyses of the actual observed CALIOP–ALADIN collocated dataset containing ~78 000 pairs of nighttime SR(z) profiles revealed the following points: (a) the values of SR(z) agree well up to ~3 km height. (b) The CF(z) profiles show agreement below ~3 km, where ~80 % of the clouds detected by CALIOP are detected by ALADIN as expected from the numerical experiment. (c) Above this height, the CDA$_{norm}^{obs}$(z) reduces to ~50 %. (d) On average, better sensitivity to lower clouds skews ALADIN's cloud peak height in pairs of ALADIN–CALIOP profiles by ~0.5±0.6 km downwards, but this effect does not alter the heights of polar stratospheric clouds and high tropical clouds thanks to their strong backscatter signals. (e) The temporal evolution of the observed CDA$_{norm}^{obs}$(z) does not reveal any statistically significant change during the considered period. This indicates that the instrument-related issues in ALADIN L0/L1 have been mitigated, at least down to the uncertainties of the following CDA$_{norm}^{obs}$(z) values: 68±12 %, 55±14 %, 34±14 %, 39±13 %, and 42±14 % estimated at 0.75, 2.25, 6.75, 8.75, and 10.25 km, respectively.

54 ENVIRONMENTAL SCIENCES↗
DOE OSTI2025

Mesoscale Cellular Convection Detection and Classification Using Convolutional Neural Networks: Insights From Long-Term Observations at ARM Eastern North Atlantic Site

Marine boundary layer clouds are crucial in Earth's climate system. They frequently manifest as closed or open cell mesoscale cellular convection (MCC). MCC clouds are challenging to represent accurately in current climate models, highlighting the need for detailed observational data sets and in-depth analyses. This study utilizes over 8 years of observations from the U.S. Department of Energy (DOE) Atmospheric Radiation Measurement (ARM) User Facility Eastern North Atlantic (ENA) site at Graciosa Island, Azores, to investigate these clouds. We first apply a convolutional neural network with a U-Net architecture to classify open and closed cells, marking the first application of such an approach for automatically detecting MCC patterns from ground-based radar measurements. This method addresses some observational gaps in satellite data related to low temporal resolution, nighttime challenges, and limited vertical structure capture. The analysis of the MCC cases shows clear differences between closed and open MCCs: Closed MCC clouds are characterized by lower cloud tops and bases, shallower cloud geometrical depth, weaker horizontal wind speeds, stronger atmospheric stability, and a more homogeneous liquid water path than open MCCs. Finally, we demonstrate two potential applications of our radar-based MCC classifications: (a) facilitating the investigation of aerosol-cloud interactions and (b) exploring meteorological factors along with MCC's evolution by integrating satellite imagery and back-trajectory analysis. The identified MCC cases offer a valuable resource for the scientific community to study MCC processes further and improve climate model accuracy.

54 ENVIRONMENTAL SCIENCES↗
DOE OSTI2024

Detecting Important Drivers of Gridded Population Modeling With Machine Learning

High-resolution population datasets have been lever-aged across a broad swath of domains, such as climate change, public policy, humanitarian aid, and rescue operations, among others. Machine learning methods were adopted to generate high-resolution or gridded population estimates by using various geospatial input features such as buildings, roads, and nighttime lights. In this study, we evaluate the importance of population features using Random Forest models across three levels of analysis, utilizing permutation measures. Our research aims to address key questions to enhance our understanding of high-resolution population modeling, such as: Are certain features globally (10 countries collectively) more important than others? Do optimal features vary by country? Within each country, do feature importance differ across administrative units? What similarities exist in feature importance at the global, country, and administrative unit levels? To answer these questions, we leverage the Kneedle algorithm to automate the selection of optimum features. We find that there are patterns displayed by features across spatial boundaries, evidenced by the same feature being the most important indicator of population across 7 of the 10 countries modeled. Our findings indicate that while important features may vary across geographies, certain features consistently hold greater importance than others agnostic of geography.

Lebakula, Viswadeep [ORNL] (ORCID:0000000152935914↗
DOE OSTI2023

Diel, seasonal, and inter-annual variation in carbon dioxide effluxes from lakes and reservoirs

Abstract Accounting for temporal changes in carbon dioxide (CO 2 ) effluxes from freshwaters remains a challenge for global and regional carbon budgets. Here, we synthesize 171 site-months of flux measurements of CO 2 based on the eddy covariance method from 13 lakes and reservoirs in the Northern Hemisphere, and quantify dynamics at multiple temporal scales. We found pronounced sub-annual variability in CO 2 flux at all sites. By accounting for diel variation, only 11% of site-months were net daily sinks of CO 2 . Annual CO 2 emissions had an average of 25% (range 3%–58%) interannual variation. Similar to studies on streams, nighttime emissions regularly exceeded daytime emissions. Biophysical regulations of CO 2 flux variability were delineated through mutual information analysis. Sample analysis of CO 2 fluxes indicate the importance of continuous measurements. Better characterization of short- and long-term variability is necessary to understand and improve detection of temporal changes of CO 2 fluxes in response to natural and anthropogenic drivers. Our results indicate that existing global lake carbon budgets relying primarily on daytime measurements yield underestimates of net emissions.

54 ENVIRONMENTAL SCIENCES↗
DOE OSTI2024

Identification of surface urban heat versus cool islands for arid cities depends on the choice of urban and rural definitions

The urban heat island (UHI) effect in arid cities can be small or even negative, the latter known as the urban cool island (UCI) effect. Differences in defining urban and rural areas can introduce uncertainties in detecting UHI or UCI, especially when the UHI signal is small. Here, we compared the surface UHI intensity (SUHII) estimated by a dozen different methods (with multiple urban and/or rural definitions) across 104 arid cities globally, providing a comprehensive evaluation of the uncertainty in SUHII estimates. Results show that the absolute difference in annual average SUHII (ΔSUHII) among methods exceeded 1°C in about half of the arid cities during both daytime and nighttime. Further, the overall annual mean ΔSUHII for all arid cities was 1.35°C during daytime and 1.03 °C at night. The uncertainty arising from simultaneous variations in urban and rural definitions was generally higher than that resulting from their individual changes. It was observed that, with varying definitions of urban and rural areas, nearly 50% of arid cities experienced a sign reversal in daytime SUHII estimates, while approximately 15% exhibited a sign reversal in nighttime SUHII. Variations in urban-rural differences in surface properties, such as vegetation index and albedo, due to differing urban and rural definitions, contributed strongly to the observed SUHII uncertainties. Overall, our results offer new insights into the ongoing debate on heat and cold islands in arid cities, emphasizing a critical need to standardize SUHII estimation frameworks.

54 ENVIRONMENTAL SCIENCES↗
DOE OSTI2024

Four Years of Atmospheric Boundary Layer Height Retrievals Using COSMIC-2 Satellite Data

This work aimed to study the atmospheric boundary layer height (ABLH) from COSMIC-2 refractivity data, endeavoring to refine existing ABLH detection algorithms and scrutinize the resulting spatial and seasonal distributions. Through validation analyses involving different ground-based methodologies (involving data from lidar, ceilometer, microwave radiometers, and radiosondes), the optimal ABLH determination relied on identifying the lowest refractivity gradient negative peak with a magnitude at least $τ$% times the minimum refractivity gradient magnitude, where $τ$ is a fitting parameter representing the minimum peak strength relative to the absolute minimum refractivity gradient. Different $τ$ values were derived accounting for the moment of the day (daytime, nighttime, or sunrise/sunset) and the underlying surface (land or sea). Results show discernible relations between ABLH and various features, notably, the land cover and latitude. On average, ABLH is higher over oceans (≈1.5 km), but extreme values (maximums > 2.5 km, and minimums < 1 km) are reached over intertropical lands. Variability is generally subtle over oceans, whereas seasonality and daily evolution are pronounced over continents, with higher ABLHs during daytime and local wintertime (summertime) in intertropical (middle) latitudes.

54 ENVIRONMENTAL SCIENCES↗
DOE OSTI2022

Detecting impacts of surface development near weather stations since 1895 in the San Joaquin Valley of California

Abstract Temperature readings observed at surface weather stations have been used for detecting changes in climate due to their long period of observations. The most common temperature metrics recorded are the daily maximum (TMax) and minimum (TMin) extremes. Unfortunately, influences besides background climate variations impact these measurements such as changes in (1) instruments, (2) location, (3) time of observation, and (4) the surrounding artifacts of human civilization (buildings, farms, streets, etc.) Quantifying (4) is difficult because the surrounding infrastructure, unique to each site, often changes slowly and variably and is thus resistant to general algorithms for adjustment. We explore a direct method of detecting this impact by comparing a single station that experienced significant development from 1895 to 2019, and especially since 1970, relative to several other stations with lesser degrees of such development (after adjustments for the (1) to (3) are applied). The target station is Fresno, California (metro population ~ 15,000 in 1900 and ~ 1 million in 2019) situated on the eastern side of the broad, flat San Joaquin Valley in which several other stations reside. A unique component of this study is the use of pentad (5-day averages) as the test metric. Results indicate that Fresno experienced + 0.4 °C decade −1 more nighttime warming (TMin) since 1970 than its neighbors—a time when population grew almost 300%. There was little difference seen in TMax trends between Fresno and non-Fresno stations since 1895 with TMax trends being near zero. A case is made for the use of TMax as the preferred climate metric relative to TMin for a variety of physical reasons. Additionally, temperatures measured at systematic times of the day (i.e., hourly) show promise as climate indicators as compared with TMax and especially TMin (and thus TAvg) due to several complicating factors involved with daily high and low measurements.

54 ENVIRONMENTAL SCIENCES↗
DOE OSTI2025

Enhanced Boundary Layer Height Detection Using Ceilometer, Surface Meteorology, and Radiation Products With a Random Forest Ensemble Method

This study develops and evaluates a Random Forest (RF) model for estimating planetary boundary layer height (PBLH) using 9 years of data from the Atmospheric Radiation Measurement Southern Great Plains (ARM SGP) user facility, with potential application in the NOAA Surface Radiation (SURFRAD) Network. The model integrates ceilometer, surface meteorology, and radiation measurements, and is trained using thermodynamic PBLH estimates derived from radiosondes. This approach aims to bridge gaps between aerosol-based and thermodynamic-based PBLH estimates. The RF model outperformed traditional methods during daytime and better captured transition periods, demonstrating improved accuracy and robustness. At ARM SGP, it showed a substantial reduction in both bias and RMSE, with a bias near zero (−4.9 m) compared with traditional Haar Wavelet (HW) (70.9 m) and Vaisala BL-View software (124.1 m), and an RMSE of 303.2 m, lower than both BL-View (566.9 m) and HW (404.6 m). During daytime hours, RF consistently outperformed both alternatives, maintaining lower bias and RMSE across all periods. At a second evaluation site, RF achieved the lowest overall RMSE (323.7 m), similar to HW (326.4 m) and significantly better than BL-View (738.3 m). However, all models showed reduced accuracy under stable nighttime conditions, limiting the reliability of PBLH estimates. Key predictors for the model included the lifting condensation level height (LCLH), aerosol gradients, and month for seasonal variability. The study underscores the potential of integrating machine learning with multiple data sets such as surface energy and thermodynamic data to advance PBLH estimation.

boundary layer height↗
DOE OSTI2022

Detection of surface water temperature variations of Mongolian lakes benefiting from the spatially and temporally gap-filled MODIS data

Lakes provide critical water resources for human activities and ecosystems, particularly in the Mongolian Plateau (MP), which is characterized by a dry climate and a harsh environment. As a region that is sensitive to anthropogenic warming, tracking lake surface water temperature (LSWT) changes in Mongolian lakes is crucial for understanding the consequences of a warming climate on lake ecosystems. However, the long-term monitoring of LSWT is restricted by the spatiotemporal gaps in the raw imagery of remote sensing-based land surface temperature (LST), e.g., the commonly used Moderate Resolution Imaging Spectroradiometer (MODIS) LST products. This study applied an improved gap-filling method by utilizing the discrete cosine transform-based penalized least squares (DCT-PLS) strategy in the spatial domain combined with the linear interpolation (LI) algorithm in the temporal domain. The method was applied to fill gaps in the LSWT imagery of 12 representative lakes across MP. The randomly sampled high-quality MODIS LSWT values in the spatial and temporal domains were excavated as false data gaps and considered “virtual true” validation datasets. The spatial validation results showed that the estimated LSWT for all the lake cases were comparable with the “virtual true” LSWT values, with the average values of the coefficient of determination, mean absolute error, mean square error, and root mean square error being 0.98, 0.38 °C, 0.45 °C, and 0.59 °C, respectively. Meanwhile, the error of nighttime LSWT results was relatively lower than that of daytime LSWT. For temporal interpolation validation, the LI algorithm exhibited relatively better performance and could more objectively indicate the variation in LSWT. Benefiting from the spatially and temporally well-constrained data, we analyzed the interannual and intra-annual change characteristics of the LSWTs of the 12 lakes. The long-term variations of annual and seasonal mean LSWTs in the 12 selected lakes exhibited no evident trends in 2000–2020, while presented apparent interannual fluctuations. The slight changes in the average LSWTs of the 12 selected lakes were in excellent synchronization with the surrounding LST derived from the reanalysis datasets, confirming the widely reported phenomenon of “global warming hiatus” that occurred in the early 21st century. This study improves the understanding of the LSWT variations in Mongolian lakes in response to global climate change. It has the potential to provide an effective approach for monitoring LSWT changes in other large-scale studies.

54 ENVIRONMENTAL SCIENCES↗
DOE OSTI2022

Profile Images and Annotations for Vehicle Re-identification Algorithms (PRIMAVERA)

This dataset contains 636,246 profile images of vehicles representing 13,963 unique vehicles. The data was collected by a set of roadside sensors over the course of three years. Each time a vehicle passed by one of the sensors, a series of images was collected. The images were processed to detect and localize each vehicle, and a license plate reader collocated with the sensor was used to provide a unique ID for the vehicle. Actual license plate numbers have been obfuscated by replacing with an arbitrary numerical ID for each vehicle. After localizing the vehicle in each image, the original RGB image was rotated, scaled, and shifted to produce a new RGB image of size 234x234 pixels such that the outermost two wheels are located at predetermined pixel locations in the image. In this way, all vehicle images are aligned to one another. This registration process occasionally results in a portion of certain vehicles being cutoff at the edges of the image. The dataset has been partitioned into two sets called training and validation. The two partitions no common vehicles, i.e., a vehicle present in one partition is guaranteed not to be present in the other. In this way, an algorithm can be validated against a set of new vehicles that were not seen during the training process. The training set contains 543,926 images from 64,440 vehicle passes representing 11,918 unique vehicles, while the validation set contains 92,320 images from 10,991 vehicle passes representing 2,045 unique vehicles. Vehicle images are organized by directories corresponding to unique vehicles. The file naming scheme is as follows: veh_{vehID}_tr_{passID}_{frameID}_{elevation}_{timeofday}.jpg where {vehID} is the vehicle ID (unique across the entire dataset), {passID} is an identifier for each tracked vehicle pass (unique across the entire dataset), {frameID} is the index of the frame within the given vehicle pass starting at 0, {elevation} is a two-letter string indicating whether the sensor was elevated (el) or at ground-level (gl), and {timeofday} is a two-letter string indicating whether the image was captured during daytime (dt) or nighttime (nt).

image↗
DOE OSTI2026

Two-Stage Wildlife Event Classification for Edge Deployment

Camera-based wildlife monitoring is often overwhelmed by non-target triggers and slowed by manual review or cloud-dependent inference, which can prevent timely intervention for high stakes human–wildlife conflicts. Our key contribution is a deployable, fully offline edge vision sensor that achieves near-real-time, highly accurate wildlife event classification by combining detector-based empty-image suppression with a lightweight classifier trained with a staged transfer-learning curriculum. Specifically, Stage 1 uses a pretrained You Only Look Once (YOLO)-family detector for permissive animal localization and empty-trigger suppression, and Stage 2 uses a lightweight EfficientNet-based binary classifier to confirm puma on detector crops and gate downstream actions. Our design is robust to low-quality nighttime monochrome imagery (motion blur, low contrast, illumination artifacts, and partial-body captures) and operates using commercially available components in connectivity-limited settings. In field deployments running since May 2025, end-to-end latency from camera trigger to action command is approximately 4 s. Ablation studies using a dataset of labeled wildlife images (pumas, not pumas) show that the two-stage approach substantially reduces false alarms in identifying pumas relative to a full-image classifier while maintaining high recall. On the held-out test set (N = 1434 events), the proposed two-stage cascade achieves precision 0.983, recall 0.975, F1 0.979, accuracy 0.986, and balanced accuracy 0.983, with only 8 false positives and 12 false negatives. The system can be easily adapted for other species, as demonstrated by rapid retraining of the second stage to classify ringtails. Downstream responses (e.g., notifications and optional audio/light outputs) provide flexible actuation capabilities that can be configured to support intervention.

58 GEOSCIENCES↗