Macao-ebird: A Curated Dataset for Artificial-Intelligence-Powered Bird Surveillance and Conservation in Macao

Abstract

Artificial intelligence (AI) currently exhibits considerable potential within the realm of biodiversity conservation. However, high-quality regionally customized datasets remain scarce, particularly within urban environments. The existing large-scale bird image datasets often lack a dedicated focus on endangered species endemic to specific geographic regions, as well as a nuanced consideration of the complex interplay between urban and natural environmental contexts. Therefore, this paper introduces Macao-ebird, a novel dataset designed to advance AI-driven recognition and conservation of endangered bird species in Macao. The dataset comprises two subsets: (1) Macao-ebird-cls, a classification dataset with 7341 images covering 24 bird species, emphasizing endangered and vulnerable species native to Macao; and (2) Macao-ebird-det, an object detection dataset generated through AI-agent-assisted labeling using grounding DETR with improved denoising anchor boxes (DINO), significantly reducing manual annotation effort while maintaining high-quality bounding-box annotations. We validate the dataset’s utility through baseline experiments with the You Only Look Once (YOLO) v8–v12 series, achieving a mean average precision (mAP50) of up to 0.984. Macao-ebird addresses critical gaps in the existing datasets by focusing on region-specific endangered species and complex urban–natural environments, providing a benchmark for AI applications in avian conservation.

1. Introduction

Earth’s biodiversity is facing an unprecedented crisis, with species extinction rates accelerating due to habitat loss, climate change, pollution, and overexploitation [1,2]. Birds, as key indicators of environmental health, are particularly vulnerable to these threats, with many species experiencing significant population declines and range contractions [3,4]. Understanding and monitoring bird species and populations are crucial for effective conservation strategies, enabling targeted interventions to protect vulnerable species and their habitats [5]. The traditional methods of bird identification and monitoring often rely on expert ornithologists, which can be time-consuming, costly, and limited in geographic scope [6]. Citizen science initiatives, where volunteers contribute data on bird sightings, have become increasingly valuable, but they can be subject to biases and require rigorous quality control [7,8].

Artificial intelligence (AI), particularly computer vision techniques, offer a promising alternative for automated bird identification and monitoring [9,10]. Deep learning (DL) models, such as convolutional neural networks (CNNs) and object detection algorithms, have demonstrated remarkable accuracy in identifying bird species from images and videos [11,12]. However, the success of these AI-based methods critically depends on the availability of large, high-quality labeled datasets [13]. The existing bird datasets, such as Caltech-UCSD Birds 200 (CUB-200) [14] and the iNaturalist dataset [15], are valuable resources. Nevertheless, they may not adequately represent the specific bird species found in particular geographic regions, especially those with unique ecological characteristics. Furthermore, these datasets often lack sufficient data on endangered species, which are a priority for conservation efforts.

Macao, despite its small size and high population density, is an important stopover point for migratory birds and supports a surprisingly diverse range of avian species [16,17]. The unique blend of urban and natural environments in Macao creates both opportunities and challenges for bird conservation. However, information about Macao’s avifauna is scattered, not easily accessible, and often lacks the detail needed for effective conservation planning. In addition, the urban landscape presents a significant threat, impacting habitats and presenting obstacles for migrating species [18].

To address these challenges, we introduce Macao-ebird, a novel dataset specifically designed for bird species identification in Macao, with a particular focus on endangered species. Macao-ebird aims to provide a valuable resource for researchers and practitioners interested in developing AI-powered tools for avian conservation in this unique urban environment. This paper makes the following contributions:

1.

Macao-ebird-cls: A classification dataset containing 7341 images of 24 bird species by data crawling, emphasizing endangered and vulnerable species in Macao. This dataset is specifically curated to reflect the local avifauna and conservation priorities.

2.

Macao-ebird-det: An object detection dataset created using AI-agent-assisted labeling via grounding DINO (DETR with improved denoising anchor boxes) [19], providing bounding-box annotations for bird instances. This innovative approach significantly reduces the manual effort required for creating a large-scale detection dataset.

3.

Baseline Experiments: A demonstration of Macao-ebird-det’s utility through baseline experiments using You Only Look Once (YOLO) v8–v12 [20,21,22,23,24] for bird species detection. These experiments provide a benchmark for future research and development using the dataset.

2. Related Works

2.1. Bird Datasets

Publicly available datasets have been instrumental in advancing avian research. The CUB-200-2011 dataset [14], containing 11,788 images of 200 North American bird species, became a benchmark for fine-grained classification. Similarly, iNaturalist 2017 [25] and BirdCLEF [26] extended coverage to global species with audio–visual data. However, these datasets lack annotations for endangered birds in specific geographic regions. For instance, Macao’s critically endangered Black-faced Spoonbill (Platalea minor) is under-represented in the existing resources. Recent efforts like the iNaturalist 2021 Birds Dataset [27] partially address this gap by crowdsourcing images, but domain-specific challenges (e.g., small sample sizes and occlusions) remain unresolved. To the best of our knowledge, there is no bird dataset specifically for Macao, especially for endangered or nationally protected birds.

2.2. AI-Agent-Assisted Labeling

The existing assisted labeling methods, such as active learning [28] and semi-supervised techniques [29], reduce manual effort but remain constrained by predefined taxonomies, failing to detect novel or rare species (open-vocabulary detection). Interactive tools like the segment anything model (SAM) [30] require heavy human input for fine-grained species differentiation while lacking semantic alignment between visual features and ecological knowledge. Grounding DINO addresses these gaps through text-guided open-set detection. By aligning image regions with text embeddings (e.g., “find the bird Platalea minor.”), it enables zero-shot localization of unseen species without bounding-box annotations, significantly reducing dependency on labeled data—critical for endangered avian conservation. By leveraging text-guided localization, it reduces manual bounding-box annotation efforts, cutting the labeling costs by 50–70% [31]. These advancements effectively address the core challenge of scarce annotated data for endangered avian monitoring and provide a scalable solution for biodiversity conservation.

2.3. Bird Species Classification and Detection

The traditional methods for bird species recognition relied on handcrafted features such as color histograms and texture descriptors (e.g., Gabor filters) [32] but struggled with lighting variations, cluttered backgrounds, and taxonomic similarities. DL then revolutionized the field: CNNs like ResNet [33] and EfficientNet [34] achieved 85–98% accuracy on large datasets by learning hierarchical visual patterns, while YOLO-based architectures [35,36,37,38] enabled real-time detection of small objects [39,40].

Recent advances addressed data scarcity through transfer learning. Vision transformers (ViTs) like data-efficient image transformers (DeiT) [41] and pre-trained models like BirdNET [42] leveraged domain knowledge from ImageNet to improve low-data performance. Data augmentation techniques such as CutMix [43], MixUp [44], and generative adversarial networks (GANs) generated images [45] that partially mitigated the data limitations but required careful annotation. Multimodal approaches combined acoustic and visual data: BirdCLEF challenges [46] integrated sound for urban bird recognition, while studies demonstrated the potential of audio–visual sensors to detect endangered species like Platalea minor [47,48]. However, while global datasets like Birdsnap [49] and CUB-200-2011 provide broad taxonomic coverage, there is a lack of Macao-specific datasets of endemic or protected species for training the DL models. Less than 5% of the existing studies focus on South China’s avian fauna, with none tailored to Macao’s unique urban ecosystems [50]. This data scarcity not only limits the model accuracy in urbanized regions like Macao but also hinders practical applications for critical conservation tasks—such as monitoring habitat fragmentation effects on migratory birds [51].

3. Dataset Creation

The Macao-ebird dataset was constructed using a combination of publicly available online resources and the AI-agent-assisted labeling process. We prioritized species listed as endangered or vulnerable by The Catalogue of Birds in the Cotai Ecological Zone [52] and Report on the State of Macao’s Environment [53], which is published from Macao SAR Government Environmental Protection Bureau. These species belong to China Red Data Book of Endangered Animals [54] or China Protection Animal Classes I and II [55]. This selection aimed to create a dataset focused on species most in need of monitoring and conservation efforts.

3.1. Macao-ebird-cls

3.1.1. Image Acquisition via Web Scraping

Ornithology and bird conservation efforts are increasingly reliant on digital resources that facilitate data collection, identification, and comprehensive species understanding. Images for the Macao-ebird-cls dataset were initially acquired using web scraping methodologies from two primary sources. We leveraged global bird observation and recording platforms, specifically eBird (Cornell Lab of Ornithology) and Observation.org. Images from these platforms generally came with initial species labels assigned by the users who submitted the observation records. We also utilized Baidu Image Search. For this source, candidate images were retrieved by executing search queries based on the scientific or common names of the target bird species. The initial presumptive labels for these images were thus inferred from the search terms or accompanying image metadata and captions.

For citizen science platforms such as eBird and Observation.org, data (including images) are typically contributed by users under specific terms that often permit use for research and non-commercial purposes, frequently adhering to Creative Commons licenses. We have complied with the terms of use of these platforms. For images obtained through general web searches (e.g., Baidu Image Search), the copyright for these images remains with the original creators. Our dataset has been compiled for non-commercial research and educational purposes, with the aim of advancing AI-based bird monitoring and conservation, which also aligns with the spirit of data sharing within the scientific and conservation communities.

3.1.2. Image Curation and Selection

After scraping the images, we implemented image curation and selection to ensure taxonomic accuracy. We initially removed duplicate images (based on file hash) and images below a certain resolution threshold (minimum dimensions of 224 × 224 pixels). Human annotators then confirmed species identity, assessed image quality (clarity, lighting, and composition), removed occluded images (e.g., images where branches obscured the bird), and excluded images of captive birds. It is important to note that the dataset only retains images that contain only one instance (i.e., one image corresponds to one instance). The process reduces the number of new categories formed by bird combinations and avoids ambiguity caused by subsequent mixed-label annotations.

3.1.3. Dataset Statistics

Table 1. The instance numbers of Macao-ebird-cls dataset.

ID	Bird Species	Number of Instances
0	Accipiter nisus	323
1	Arenaria interpres	251
2	Calidris falcinellus	260
3	Calidris tenuirostris	227
4	Calliope calliope	389
5	Centropus sinensis	281
6	Circus spilonotus	412
7	Egetta eulophotes	330
8	Egretta sacra	379
9	Elanus caeruleus	304
10	Falco amurensis	276
11	Falco tinnunculus	271
12	Garrulax canorus	280
13	Halcyon smyrnensis	340
14	Hydrophasianus chirurgus	251
15	Leiothrix argentauris	358
16	Leiothrix lutea	386
17	Limnodromus semipalmatus	245
18	Merops philippinus	316
19	Milvus migrans	389
20	Numenius arquata	257
21	Pandion haliaetus	283
22	Platalea leucorodia	266
23	Platalea minor	267
	Total instances	7341

shows the number of instances in the Macao-ebird-cls dataset. We obtained 7341 images across 24 bird species. Overall, the number of images of most birds is between 200 and 400, showing that the order of magnitude of image records is relatively concentrated. The bird with the largest number is Circus spilonotus (412 instances), while the bird with the smallest number is Halcyon smyrnensis (340 instances).

Figure 1 shows samples of 24 bird species of the Macao-ebird-cls dataset. The illustrations cover the life scenes of most endangered birds (shallow water, mudflats, branches, grass, sky, etc.). Their sizes and locations are also different, ensuring the diversity of samples.

3.2. Macao-ebird-det

The images used for the detection dataset were the images collected for the classification dataset. We prioritized images that contained birds in complex environments to create a challenging and realistic detection task.

3.2.1. AI-Agent-Assisted Labeling

Grounding DINO is an open-set object detector that bridges the gap between vision and language by enabling zero-shot object detection capabilities. Unlike traditional object detectors trained on fixed categories, grounding DINO can detect objects specified through natural language prompts, even those it has not seen during training. This is achieved by leveraging a transformer-based architecture trained on a large dataset of image–text pairs, enabling it to understand the semantic relationships between objects and their textual descriptions. It has become a popular choice for tasks requiring flexible object detection and semantic understanding of visual scenes.

Figure 2 is the multi-stage depiction of AI-agent-assisted labeling via grounding DINO. It begins with an input image and a natural language text prompt. The image is processed by a CNN-based image backbone (blue), extracting visual features, while the text prompt (e.g., “find the bird: Accipiter nisus”) is encoded by a transformer-based text backbone (green) into semantic embeddings. Subsequently, a feature enhancer (brown) refines both feature sets. Crucially, the language-guided query selection (brown) module employs cross-modal attention, enabling the text embeddings to guide the selection of relevant visual features, resulting in object queries. These queries are then fed into a decoder (yellow), which predicts object bounding boxes and confidence scores. Finally, a post-processing (gray) step refines the predictions via non-maximum suppression (NMS) and confidence thresholding, resulting in the output: the input image with overlaid bounding boxes accurately localizing the objects specified in the text prompt.

3.2.2. Manual Verification and Refinement

The initial bounding-box annotations generated by grounding DINO were manually reviewed and refined by human annotators to ensure accuracy. This included correcting bounding-box positions, adding missing annotations (e.g., for birds that were partially obscured), and removing false positives (e.g., objects that were mistakenly identified as birds). We used a specialized annotation tool (e.g., LabelImg) to facilitate this process. We map the generated labels and bounding-box positions to the corresponding input images in batches and check them, including no labeled information or wrong labels and bounding-box positions.

The final instance counts for the Macao-ebird-det dataset are detailed in

Table 2. The final instance counts for the Macao-ebird-det dataset by AI agent using grounding DINO.

Dataset	Number of Instances
Macao-ebird-cls (the original dataset)	7341
Labels generated failed	54
Macao-ebird-det (successfully labeled)	7287
Labeling success rate	99.18%

. From an initial set of 7341 input images encompassing 24 bird species (Macao-ebird-cls dataset), label and annotation generation was unsuccessful for 54 images. Consequently, 7287 images (instances) were successfully labeled, representing a labeling success rate of 99.18%. The methodology, which utilized bird folder categories from the classification dataset as input prompts, significantly reduced the likelihood of generating incorrect labels and bounding-box positions.

3.2.3. Dataset Statistics

Figure 3 includes the annotated instance numbers of Macao-ebird-det dataset with a total of 7287 images across 24 bird species. Image counts ranged from a minimum of 227 for Limnodromus semipalmatus and Calidris tenuirostris to a maximum of 411 for Circus spilonotus. The mean number of images per species was 307.96, with a median of 281.50. The dataset exhibits uniform class distribution with minimal outliers and negligible imbalance, demonstrating robust sampling methodology.

Figure 4 shows two 2D histograms (density scatter plots) visualizing characteristics of bounding-box distribution from Macao-ebird-det dataset.

The left plot shows the distribution of the normalized center coordinates (x, y) of the bounding boxes. There is a strong concentration of data points (indicated by darker blue squares) around the center of the plot ($x\approx 0.5$, $y\approx 0.5$). The density decreases radially outwards from this central point. This indicates that the objects detected in the dataset are predominantly located near the center of the image frames. Objects appearing near the edges of the images are less common.

This right plot shows the relationship between the normalized width and height of the bounding boxes. A clear positive correlation exists between width and height, visible as a diagonal trend extending from the bottom-left towards the top-right. The highest density of points are concentrated in the region where both width and height are relatively small (e.g., typically less than 0.4). The density decreases as either width or height (or both) increase. This signifies that width and height tend to scale together, which is expected as they represent the dimensions of physical objects. The plot strongly indicates that the dataset contains a large number of relatively small objects (small width and small height), while larger objects are less frequent. The correlation suggests a range of aspect ratios but generally avoids extreme cases like very wide but short boxes or very tall but thin boxes.

In summary, the visualizations suggest the Macao-ebird-det dataset primarily contains objects centered within the image frames and that most of these objects are relatively small in size, with width and height showing a positive correlation.

Figure 5 shows samples of annotated images of 24 bird species from the Macao-ebird-det dataset. As demonstrated, the birds in the samples are accurately detected. Bounding boxes effectively encompass their bodies and class labels correctly match the detected species. Even when birds exhibit diverse poses or are situated in complex environmental backgrounds, the AI agent still performs well in detection tasks.

4. Baseline Experiments

4.1. Training Details

We utilized YOLOv8-v12 for our baseline experiments. Specifically, we selected the “n” and “s” versions of the models in each series for training and evaluation. These versions have small computational complexity and low requirements on device computing power. They can be easily deployed to edge devices for real-time deployment and reasoning in the wild.

Baseline experiments are conducted in DL framework torch 1.12.1 with cu116 and trained on Graphics Processing Unit (GPU) NVIDIA GeForce 1080. The dataset The Macao-ebird-det dataset is divided into training, validation, and testing datasets in a 7:2:1 ratio. Specifically, for each bird species within the dataset, its corresponding images were randomly partitioned into 70% for the training dataset, 20% for the validation dataset, and 10% for the test dataset. The purpose of employing this stratified sampling strategy is to ensure class balance is maintained across the training, validation, and test datasets. The training phase utilizes transfer learning and data augmentation. Dataset images, after being loaded into memory during the training process, are all resized to 640x640 to ensure that the input image dimensions match those used during model training. The optimizer uses the auto mode with an initial learning rate of 0.01. dynamically adjusting the configuration during training. The batch size of a single GPU is 16, and the training round is 200 epochs. Giga floating point of operations (GFLOPs), precision, recall, and mean average precision (mAP50 and mAP50-90) are applied to evaluate the detection performance.

4.2. Discussion

Table 3. The performance evaluation of baseline experiments. All results are obtained using 640 × 640 inputs.

Model	#Param.	FLOPs	Precision	Recall	mAP50	mAP50-95	Speed 1
	(M)	(G)					(ms)
YOLOv8n	3.01	8.1	0.951	0.943	0.975	0.939	2.3
YOLOv8s	11.13	28.5	0.966	0.949	0.98	0.95	5.1
YOLOv9t	1.98	7.6	0.954	0.938	0.977	0.945	3.3
YOLOv9s	7.18	26.8	0.972	0.96	0.984	0.958	6.6
YOLOv10n	2.73	8.3	0.957	0.941	0.977	0.945	3.1
YOLOv10s	8.05	24.5	0.973	0.948	0.98	0.955	6.5
YOLOv11n	2.59	6.3	0.951	0.938	0.972	0.938	2.4
YOLOv11s	9.42	21.4	0.961	0.954	0.977	0.953	5.4
YOLOv12n	2.56	6.3	0.942	0.936	0.972	0.936	4
YOLOv12s	9.24	21.3	0.97	0.953	0.982	0.957	8.2

¹ Inference speed metrics refer to the model inference speed excluding pre-processing and post-processing. Speeds are averaged over Macao-ebird-det validation dataset using an NVIDIA GeForce GTX 1080 instance.

includes the performance evaluation of the baseline experiments. We analyze and discuss the baseline experiments from four perspectives: annotation consistency, efficiency and inference speed, data distribution balance, and challenging case coverage.

4.2.1. Annotation Consistency Verification

The lightweight models (e.g., YOLOv11n and YOLOv12n) exhibit stable mAP50 values (0.936–0.972), confirming minimal annotation noise. Cross-model consistency is further validated by all the models achieving mAP50 > 0.972 (peaking at 0.984). This demonstrates precise bounding-box annotations with minimal localization errors at an Intersection over Union (IoU) threshold of 50%.

4.2.2. Efficiency and Inference Speed Analysis

Figure 6 reveals the expected trade-offs between model size, computational cost (FLOPs), accuracy (mAP), and inference speed based on the corresponding results in Table 3. Within each version family (v8, v9, v10, v11, and v12), the “s” models consistently have more parameters and higher FLOPs than their “n”/“t” counterparts. This generally translates to higher accuracy (mAP50-95) but also significantly slower inference speeds (e.g., YOLOv8n: 2.3 ms vs. YOLOv8s: 5.1 ms; YOLOv11n: 2.4 ms vs. YOLOv11s: 5.4 ms). The “n”/“t” models offer rapid inference, with YOLOv8n (2.3 ms) and YOLOv11n (2.4 ms) being the fastest, suitable for resource-constrained environments. YOLOv12n (4 ms) is notably slower than the other nano models despite similar FLOPs to YOLOv11n.

While later series like v10/v11/v12 generally achieve high accuracy with competitive or lower FLOPs compared to earlier series (indicating good computational efficiency, e.g., YOLOv11s/12s have 21G FLOPs vs YOLOv8s’ 28.5G FLOPs for similar or better mAP50-95), their runtime efficiency (inference speed) does not show a monotonic improvement. For instance, YOLOv12s achieves the highest mAP50-95 (0.957) among the “s” models shown but has the slowest inference time (8.2 ms). YOLOv11s offers a strong balance with high accuracy (0.953 mAP50-95) and relatively fast inference (5.4 ms) for an “s” model.

4.2.3. Data Distribution Balance Verification

The dataset shows balanced class distribution, evidenced by precision–recall gaps < 0.02 (e.g., YOLOv9s: 0.972 vs. 0.96) and consistent mAP50-95 improvements across model series (e.g., YOLOv8n→v8s: 0.939→0.95). It reflects diverse scenario coverage and effective distinction of the model capacities.

4.2.4. Challenging Case Coverage Verification

A significant mAP50-95 variation (maximum difference is 0.022) highlights the dataset’s inclusion of challenging cases (such as small/dense, edge, or truncated bird samples), driving high-performance models (e.g., YOLOv9s) to adopt complex architectures for stability under stricter IoU thresholds. For YOLOv10s, precision (0.973) is significantly higher than recall (0.948). The possible reason is that the model filters out challenging samples more aggressively. Future work involves adding occluded/low-light samples to improve the distribution of challenging cases in the dataset.

In summary, the Macao-ebird-det dataset demonstrates excellent performance in annotation accuracy and covers a good range of scenarios, effectively supporting performance comparisons, particularly for lightweight models. The results highlight clear trade-offs between accuracy and speed. While the YOLOv9 series shows strong peak accuracy (YOLOv9s), and models like YOLOv11n/s demonstrate excellent runtime efficiency (inference speed), careful consideration of the specific balance between accuracy, computational cost (FLOPs), and real-world inference speed is necessary when selecting a model for a particular application. The data suggest that later-generation models improve computational efficiency (accuracy per FLOP), but runtime speed requires individual evaluation.

5. Applications

5.1. AI-Powered Bird Surveillance Recognition and Conservation

The Macao-ebird dataset’s primary application lies in serving as foundational data for training and evaluating artificial intelligence models, such as convolutional neural networks (CNNs) and transformers, specifically for bird classification and detection tasks. The Macao-ebird-cls subset is designed for classification algorithm development, while the Macao-ebird-det subset is tailored for object detection, enabling the assessment of various detection algorithms like the YOLO series, Faster R-CNN, and DETR in accurately locating and identifying birds within real-world scenarios.

Given its focus on Macao’s avifauna, particularly including endangered species, the dataset is especially valuable for developing automated bird identification systems targeted at Macao and ecologically similar regions like the Pearl River Delta. Deploying models trained on these data—for instance, on monitoring cameras, drones, or mobile applications—facilitates automatic real-time monitoring and identification of rare birds. This capability provides crucial technological support for conservation efforts, including population counting, habitat analysis, and threat alerting, thereby actively promoting the effective protection of endangered bird species.

5.2. Citizen Science and Environmental Education

Containing over 7300 images with species labels, the Macao-ebird dataset constitutes a rich visual library suitable for public science and environmental education. These images can be utilized to develop resources such as bird identification websites, mobile applications, and educational materials like brochures or exhibits. Such tools can effectively assist the general public, students, and birdwatching enthusiasts in learning to recognize the common and rare bird species found in Macao. The detailed and representative imagery of local bird species helps to deepen citizens’ understanding of local biodiversity, particularly the status of bird populations and their habitats, consequently enhancing public awareness regarding the importance of ecological conservation and encouraging greater participation in related activities.

5.3. Algorithm Research and Benchmarking

Macao-ebird serves as a valuable resource for advancing algorithm research and establishing performance benchmarks, particularly in the domain of fine-grained visual categorization, as bird identification often requires distinguishing between visually similar species or subspecies. The dataset provides a solid foundation for testing, comparing, and evaluating the performance of various fine-grained image classification and recognition algorithms.

The baseline performance metrics reported in the paper offer a clear point of reference for future research. Researchers can leverage Macao-ebird to validate the efficacy of novel algorithms and compare different methodologies, thus fostering standardization and reproducibility within the field. Furthermore, the AI-assisted labeling approach employed (using grounding DINO) presents a case study for research into annotation efficiency and quality assessment.

6. Limitations

Although the Macao-ebird dataset prioritizes endangered and protected avian species (24 species) within Macao, it does not encompass the entirety of Macao’s avifauna, as documented in sources like The Catalogue of Birds in the Cotai Ecological Zone [53,54] (which indicates up to 174 species). This restricted species coverage limits the dataset’s direct applicability for comprehensive biodiversity assessments or the monitoring of all local bird populations. Future research aiming for a holistic evaluation of Macao’s avian ecology would necessitate dataset expansion to include a wider range of species.

Each image in the Macao-ebird dataset contains only one bird instance. This design choice simplifies image annotation and certain classification tasks. Real-world computer vision scenarios, especially in object detection and segmentation, frequently feature multiple object instances co-occurring within a single frame. Consequently, the dataset’s utility is constrained for the direct development or evaluation of algorithms specifically designed to handle multi-object detection or counting involving multiple bird instances.

The image curation process favored clear, unobstructed, and adequately sized images, leading to the exclusion of significantly occluded samples. The dataset may have a limited proportion of hard examples, such as instances with severe occlusion, high target-background similarity (camouflage), or suboptimal lighting conditions. To enhance the generalization capability and robustness of models for real-world deployment, future iterations of the dataset should incorporate greater diversity regarding challenging samples (occluded/low-light samples) to improve the distribution of challenging cases.

7. Conclusions

Macao-ebird fills a critical gap by providing a dedicated dataset for endangered bird species recognition within the unique ecological context of Macao. The dataset comprises two distinct subsets: Macao-ebird-cls and Macao-ebird-det. Macao-ebird-cls contains 7341 images across 24 species, focusing on those considered to be protected or endangered within Macao. The object detection dataset Macao-ebird-det with bounding-box annotations was created from the same images using our AI-agent-assisted labeling approach via grounding DINO. This innovative labeling method, reducing manual effort while maintaining high annotation quality, offers a significant contribution to the field. Our baseline experiments using YOLOv8-v12 demonstrate the dataset’s utility, achieving mAP50 scores of up to 0.984 and confirming minimal annotation noise. It is worth emphasizing that the dataset contains bird images of various species, scenes, and high resolutions, so it can be widely used in related research, such as bird identification, behavior analysis, and population assessment.

By bridging AI and conservation biology, Macao-ebird provides a scalable framework for protecting biodiversity in urbanized habitats. We are confident that Macao-ebird will facilitate the development of impactful solutions for protecting Macao’s unique avian biodiversity. Future work will expand the dataset to include rare species, challenging cases (e.g., small objects and dense clusters), and multimodal data (e.g., audio recordings), enhancing its applicability in ecological monitoring.