Advancing marine mammal monitoring: Large-scale UAV delphinidae datasets and robust motion tracking for group size estimation

Contents lists available at ScienceDirect Machine Learning with Applications journal homepage: www.elsevier.com/locate/mlwa Advancing marine mammal monitoring: Large-scale UAV delphinidae datasets and robust motion tracking for group size estimation Leonardo Viegas Filipea,∗, João Canelas a, Mário Vieira a, Francisco Correia da Fonseca a, André Cidb, Joana Castrob,c, Inês Machadoa aWavEC Offshore Renewables, Edifício Diogo Cão, Doca de Alcântara Norte, 1350-352, Lisbon, Portugal bAIMM - Associação para a Investigação do Meio Marinho, Rua Maestro Fred. Freitas N15-1, 1500-399, Lisbon, Portugal cMARE - Marine and Environmental Sciences Centre/ARNET - Aquatic Research Network, Laboratório Marítimo da Guia, Faculdade de Ciências da Universidade de Lisboa, Av. Nossa Senhora do Cabo, 939, 2750-374, Cascais, Portugal A R T I C L E I N F O Keywords: Artificial intelligence Marine mammal surveys UAV Object tracking Human-in-the-loop Dataset A B S T R A C T Reliable estimates of dolphin abundance are essential for conservation and impact assessment, yet manual analysis of aerial surveys is time-consuming and difficult to scale. This paper presents an end-to-end pipeline for automatic dolphin counting from unmanned aerial vehicle (UAV) video that combines modern object detection and multi-object tracking. We construct a large detection dataset of 64,705 images with 225,305 dolphin bounding boxes and a tracking dataset of 54,274 frames with 207,850 boxes and 603 unique tracks, derived from UAV line-transect surveys. Using these data, we train a YOLO11-based detector that achieves a precision of approximately 0.93 across a range of sea states. For tracking, we adopt BoT-SORT and tune its parameters with a genetic algorithm using a multi-metric objective, reducing ID fragmentation by about 29% relative to default settings. Recent YOLO-based cetacean detectors trained on UAV imagery of beluga whales report precision/recall around 0.92/0.92 for adults and 0.94/0.89 for calves, but rely on DeepSORT tracking whose MOTA remains below 0.5 and must be boosted to roughly 0.7 with post-hoc trajectory post-processing. In this context, our pipeline offers competitive detection performance, substantially larger and fully documented detection and tracking benchmarks, and GA-optimized tracking without manual post-processing. Applied to dolphin group counting, the full pipeline attains a mean absolute error of 1.24 on a held-out validation set, demonstrating that UAV-based automated counting can support robust, scalable monitoring of coastal dolphin populations. 1. Introduction Marine mammal monitoring is a critical component of conservation and management strategies, as it provides essential information on population size, distribution, and behaviour. Traditional survey methods, such as boat-based counts or manned aerial observations, are resourceintensive, weather-dependent, and can disturb target species. In recent years, drones (unmanned aerial vehicles, UAVs) have emerged as a noninvasive and cost-effective alternative for acquiring high-resolution imagery over broad spatial and temporal scales (Álvarez-González et al., 2023). In parallel, deep learning has transformed computer vision, enabling reliable detection and tracking of wildlife in large datasets. A dominant paradigm in multi-object tracking is tracking-by-detection, where each ∗Corresponding author. E-mail addresses: [email protected] (L. Viegas Filipe), [email protected] (J. Canelas), [email protected] (M. Vieira), [email protected] (F. Correia da Fonseca), [email protected] (A. Cid), [email protected] (J. Castro), [email protected] (I. Machado). frame is first processed by an object detector and detections are then linked into trajectories (Adžemović et al., 2025; Guan et al., 2025; Li et al., 2024). The success of this strategy depends directly on detection quality, as tracking continuity is built upon accurate frame-level localization. The YOLO (You Only Look Once) object detection model family has established itself as one of the most effective detectors for real-time applications, combining speed and accuracy in a unified, end-to-end model (Redmon et al., 2016). YOLO11 introduces architectural improvements that enhance efficiency and small-object detection (Jegham et al., 2025; Khanam & Hussain, 2024; Ultralytics, 2023), both crucial for identifying dolphins in drone footage. Complementing this, BoT-SORT is a state-of-the-art tracker that extends detectionbased tracking by incorporating motion cues, appearance embeddings, https://doi.org/10.1016/j.mlwa.2025.100808 Received 27 October 2025; Received in revised form 25 November 2025; Accepted 1 December 2025 Machine Learning with Applications 23 (2026) 100808 Available online 4 December 2025 2666-8270/© 2025 The Authors. Published by Elsevier Ltd. This is an open access article under the CC BY license ( http://creativecommons.org/licenses/by/4.0/ ). L. Viegas Filipe et al. and camera-motion compensation to achieve robust identity preservation (Aharon et al., 2022). The natural synergy between YOLO11 as a high-performing detector and BoT-SORT as an advanced detectionbased tracker provides a powerful foundation for monitoring dolphins from UAV imagery. Despite these advances, progress is still constrained by the limited availability of annotated datasets for model training and validation. Collecting UAV imagery of dolphins is logistically challenging and ecologically costly, and publicly available labelled datasets remain scarce. This creates barriers to developing generalizable, robust systems that can support long-term monitoring programmes rather than isolated case studies. We advance computer vision and marine ecology with an end-to-end UAV pipeline for dolphin monitoring. Our main contributions are as follows: •We curate and release a large-scale dolphin detection dataset (64,705 images with 225,305 bounding boxes) and a tracking dataset (54,274 frames with 207,850 boxes and 603 unique tracks) derived from UAV line-transect surveys, capturing a wide range of sea states and group configurations. •We train a YOLO11-based detector tailored to coastal dolphin imagery and show that it achieves high precision across variable environmental conditions, providing reliable frame-level localization for downstream tracking. •We adopt BoT-SORT as our baseline tracker and use a genetic algorithm to tune its hyperparameters with a multi-metric objective combining HOTA, IDF1, and related measures, reducing identity fragmentation compared to default settings. •We perform a detailed comparison between appearance-based and motion-only tracking variants, and we show that in this domain appearance re-identification does not improve performance over motion-only cues, informing the design of lightweight, fielddeployable systems. •We demonstrate that the resulting detection-and-tracking pipeline can be used to estimate dolphin group sizes from UAV video with low mean absolute error, illustrating the practical value of deep learning for scalable, non-invasive monitoring of coastal dolphin populations. Together, these contributions provide both new resources and concrete guidance for deploying modern computer-vision pipelines in marine mammal monitoring, and they highlight opportunities and limitations that should be considered when integrating UAV-based automated counting into conservation practice. This introduction is followed by a section dedicated to review related work done in this field (Section 2). We provide an overview of the proposed method in Section 3, detailing aspects of the different modules and overall architecture of the pipeline. After, we describe all the datasets used (Section 4) regarding acquisition and data characteristics. Next, we describe the developed annotation tool (Section 5). Model implementations are described in Section 6, starting with the training and evaluation of the detector and finishing with the multiple tracking configurations. The most relevant results are presented and discussed in Section 7. Section 8 describes limitation of our pipeline and proposes future work. Finally, Section 9 rounds up this paper presenting final conclusions. 2. Related work Deep learning has transformed computer vision, enabling reliable detection and tracking of objects in large, heterogeneous datasets. In wildlife monitoring, it has been increasingly adopted to automate tasks such as species detection, individual identification, and movement analysis. A dominant paradigm in multi-object tracking is tracking-bydetection, in which each frame is first processed by an object detector and detections are then linked into trajectories (Adžemović et al., 2025; Guan et al., 2025; Li et al., 2024). The success of this strategy depends directly on detection quality, as tracking continuity is built upon accurate frame-level localization. Several recent studies have reviewed the rapidly growing literature on deep learning for ecological and marine applications, highlighting both the potential and the limitations of current approaches (Adžemović et al., 2025; Guan et al., 2025; Li et al., 2024). Most existing work either focuses solely on detection in still images or relies on relatively small, proprietary datasets, making it difficult to assess how well models generalize across locations, environmental conditions, and survey protocols. Explicit treatment of the full pipeline from detection and tracking to biologically meaningful quantities such as group size or abundance remains comparatively rare. Recent YOLOv7-based systems have demonstrated high-precision detection of beluga whales from UAV video, achieving precision/recall around 0.92/0.92 for adults and 0.94/0.89 for calves, but rely on DeepSORT tracking whose MOTA must be boosted from 27%–48% to about 70% via post-hoc trajectory post-processing (Alsaidi et al., 2024). For marine mammals observed from UAVs, additional challenges arise. Animals are often small in the field of view, partially submerged, and embedded in a dynamic sea surface whose appearance changes with sea state, lighting, and viewing geometry. These factors complicate both detection and tracking, and they interact with survey design choices (flight altitude, camera configuration, flight paths). At the same time, UAVs offer clear operational advantages over traditional boat-based or crewed-aircraft surveys, including reduced cost and risk, higher revisit frequency, and the ability to capture stabilized video suitable for automated analysis (Álvarez-González et al., 2023). A key bottleneck across this literature is the scarcity of large, openly-described datasets that reflect the variability of real monitoring campaigns. Collecting UAV imagery of dolphins is logistically demanding and may require specialized vessels, pilots, and permits. Publicly available labelled datasets are therefore limited. The dataset introduced by Canelas et al. (2025) represents an important step in this direction, providing UAV video of coastal dolphins with frame-level annotations suitable for training and evaluating detection and tracking models under realistic survey conditions. However, even in this case, challenges remain in terms of dataset size, diversity of sea states and group configurations, and the availability of identity-consistent tracks for benchmarking multi-object tracking. Within this context, our work contributes by extending the available data resources for dolphin monitoring and by systematically evaluating a modern detection-and-tracking pipeline tailored to UAV footage of small delphinids. In contrast to studies that focus only on algorithmic performance, we explicitly link computer-vision metrics to ecological outcomes, using the pipeline to estimate dolphin group sizes from survey video and assessing errors in terms that are directly relevant to abundance estimation and conservation practice. 3. Proposed method and overall architecture Our goal is to obtain reliable dolphin group counts from UAV survey video using an automated, scalable pipeline. We adopt a tracking-bydetection strategy in which a frame-level detector produces dolphin bounding boxes and a multi-object tracker links these detections into temporally consistent trajectories. Group-level counts are then derived from the resulting tracks. Fig. 1 provides an overview of the proposed architecture. Starting from raw UAV video, we (i) extract frames and create large-scale detection and tracking datasets using a semi-supervised, human-inthe-loop annotation workflow, (ii) train a YOLO11-based detector on the detection dataset, (iii) apply this detector to new UAV frames to obtain dolphin detections, (iv) use BoT-SORT to track dolphins over time, tuning its hyperparameters with a genetic algorithm, and (v) aggregate track information into frame-level and group-level counts, which are evaluated both with standard computer-vision metrics and with ecologically meaningful error measures. Machine Learning with Applications 23 (2026) 100808 2 L. Viegas Filipe et al. Fig. 1. Overview of the proposed UAV-based dolphin counting pipeline. Raw UAV survey video is first sampled into frames and annotated using a semisupervised, human-in-the-loop workflow, producing large-scale detection and tracking datasets. A YOLO11-based detector is then trained on the detection dataset and applied to new UAV frames to produce dolphin bounding boxes. BoT-SORT, with hyperparameters tuned by a genetic algorithm, links detections across frames into trajectories. From these trajectories we obtain frame-level and group-level counts, as well as evaluation metrics such as mean absolute error (MAE), HOTA and IDF1, which quantify both counting accuracy and tracking quality. 3.1. Problem formulation Let a UAV survey produce a video sequence {𝐼𝑡}𝑇 𝑡=1 of 𝑇 frames. The primary computer-vision task is to estimate, for each frame 𝐼𝑡, the set of dolphin bounding boxes () 𝑡= {𝑏𝑡,𝑘}𝐾𝑡 𝑘=1, where 𝐾𝑡 is the number of detected dolphins in frame 𝑡. The tracking task is to assign consistent identities across frames and produce a set of trajectories  = {𝜏𝑗}𝐽 𝑗=1, where each trajectory 𝜏𝑗 is a time-ordered sequence of detections corresponding to the same individual dolphin or tightly grouped individuals. 𝐽 is the total number of trajectories in the evaluated video sequence. These trajectories are then used to estimate dolphin group sizes and group-level abundance over predefined temporal windows or survey segments. In addition to standard detection and tracking metrics (e.g. precision, recall, mean Average Precision, HOTA, IDF1), we report ecological metrics such as mean absolute error (MAE) between predicted and manually derived group counts. 3.2. Detection module The detection module processes each frame 𝐼𝑡 independently and outputs a set of dolphin detections 𝑡. We use YOLO11L as our base detector, chosen for its balance between speed and accuracy on small objects in high-resolution imagery. The network predicts bounding boxes and confidence scores, which are then filtered by non-maximum suppression to produce the final set of detections per frame. The detector is trained on the curated detection dataset described in Section 4, using a train/validation split that reflects the diversity of sea states, viewing geometries, and group configurations. Training details, including data augmentation, optimization settings, and selection of decision thresholds, are presented in Section 6. 3.3. Tracking module To obtain temporally consistent trajectories from frame-level detections, we employ BoT-SORT as our tracking backbone. BoT-SORT follows a tracking-by-detection paradigm in which detections in consecutive frames are associated using a combination of motion modelling and similarity scoring. In our configuration, the tracker maintains a set of active tracks and at each frame associates detections to existing tracks based on predicted motion and, optionally, appearance cues. Unmatched detections initialize new tracks, while tracks without recent matches are terminated. This yields a set of trajectories  that capture the movement of dolphins through the field of view. BoT-SORT performance depends on several hyperparameters governing, for example, association thresholds, the lifespan of unmatched tracks, and matching criteria. Rather than setting these heuristically, we perform a two-stage optimization: a coarse grid search followed by a genetic algorithm that refines hyperparameters using a multi-metric objective combining HOTA, IDF1 and related tracking metrics. The optimization procedure and search space are detailed in Section 6. Optionally, BoT-SORT can include a deep-learning model to aid on object Re-Identification (ReID) based on appearance cues, allowing the same individual to be recognized across frames. We performed some experiments to evaluate the benefit of appearance-based ReID, and the details of this implementation are presented in Section 6.2.3 3.4. Group counting and evaluation Once trajectories have been estimated, we derive counts at two levels: •frame-level counts, obtained by counting the number of active tracks or detections per frame after basic filtering (e.g. removal of very short tracks or low-confidence detections). •group-level counts, obtained by aggregating trajectories over survey segments corresponding to dolphin groups, following the same protocol used for manual annotations. For the computer-vision evaluation we use a dedicated tracking benchmark and report standard detection metrics: precision, recall, mean Average Precision at different Intersection over Union (IoU) thresholds; and tracking metrics: HOTA, IDF1, and related measures. For the ecological evaluation we compare predicted group sizes to expert-derived counts and report mean absolute error and related statistics, providing a direct link between algorithmic performance and monitoring accuracy. This separation between the method-level description in the present section and the implementation details in Section 6 is intended to clarify the overall architecture of the proposed solution while preserving full transparency about training protocols and hyperparameter choices. 4. Datasets 4.1. Data acquisition We used a dataset compiled by Canelas et al. (2025). Because insitu ocean surveys are costly and logistically complex, and public aerial datasets for widespread species are scarce, they focused on Delphinidae and scraped 62 videos from YouTube, Pexels, and Dailymotion across unrestricted locations, times, and conditions to build a variable dataset for training generalizable detection/tracking models. Fig. 2 shows some examples of frames extracted from these videos, illustrating scene variability. Additionally, Associação para a Investigação do Meio Marinho (AIMM) provided one video resultant from a marine mammal observation expedition. 4.2. Semi-supervised detection dataset We began by manually annotating a seed set of a few hundred images, drawing bounding boxes around dolphins. This initial dataset Machine Learning with Applications 23 (2026) 100808 3 L. Viegas Filipe et al. Fig. 2. Example of images in the dataset, illustrating scene variability. was used to train a YOLO11 detector, which was then applied to the remaining unlabelled imagery. The automatically generated predictions were reviewed and corrected by human annotators, producing an expanded, higher-quality training set. The detector was then re-trained using this refined dataset. This was repeated iteratively: YOLO11 was applied to the full dataset, its predictions corrected, and the model re-trained until the detection error stabilized at an acceptably low level, i.e., every annotator agreed with the annotations of all images. Through this cycle, the annotation burden was significantly reduced compared to fully manual labelling, while maintaining high dataset quality. Similar iterative labelling frameworks have been applied in UAV wildlife monitoring and remote sensing, confirming their suitability in data-scarce contexts (Kellenberger et al., 2018; Zhao et al., 2025). Fig. 3 illustrates this labelling method in the top blue box. This process resulted in a dataset with a total of 64 705 images with 225305 annotated boxes, giving an average of approximately 3.5 boxes per image. This dataset was partitioned in training and validation sets, using a 70–30 ratio, using, for each video, the last 30% for validation, keeping the first 70% for training, decreasing frame similarity between these two sets. 4.3. Tracking dataset The creation of high-quality tracking datasets presents additional challenges due to data scarcity. Unlike established pedestrian or vehicle tracking benchmarks, there are no large-scale, publicly available datasets for dolphin tracking from UAV imagery. This lack of annotated video data constrains the training and evaluation of multi-object trackers. To address this limitation, we adopted a model-assisted annotation workflow. The YOLO11 detector, trained on the semi-supervised dataset, was applied frame by frame, and detections were linked into tracks using the BoT-SORT algorithm. This produced initial identity assignments across sequences. Because automatic tracking can introduce errors, such as identity switches, fragmented trajectories, or false positives, a manual correction stage was performed. Annotators reviewed the results, reassigning IDs where necessary, removing false positives, and refining trajectories. This pipeline enabled the production of a reliable dolphin tracking dataset that would have been infeasible to assemble through manual annotation alone. This process is shown in Fig. 3, specifically in the bottom green box. The resultant dataset consisted of 54 274 annotated frames from 18 different videos. 207 850 boxes made 603 unique tracks. There was an average of approximately 3.8 boxes per frame. Table 1 Number of frames and total dolphin count (IDsGT) on the AIMM dataset. Seq. Frames IDsGT 0 727 14 1 407 9 2 1411 12 3 681 16 4 380 11 5 511 2 4117 64 4.4. AIMM dataset AIMM’s video was annotated for tracking using the process described in Section 4.3, and segmented into six Sequences (Seqs.) with varying number of frames, dolphin total count (Table 1), and drone height. 4.5. Sample taxonomy and background variability Taxonomically, all positive annotations in the current release correspond to delphinids; no other marine mammal or fish species are explicitly labelled. The aerial footage was collected from various internet sources, containing footage taken on nearshore and offshore waters under a range of locations, sea states and illumination conditions, with frequent whitecaps, sun glint and vessel wakes. In addition to dolphin annotations, the raw imagery contains a variety of non-dolphin objects (e.g. surfers, kayakers, fish schools, birds, watercraft, algae and rocks), which are retained as unlabelled negatives and are illustrated in Fig. 4. Implications for domain shift are discussed in Section 8. 4.6. Negative samples In the current release of these datasets we only annotate dolphins; no other marine mammal or fish species are included in this version. In practice, our aerial footage for this study does not contain other cetacean species in appreciable numbers, so the main challenge is distinguishing dolphins from sea surface clutter (whitecaps, wave crests, wakes, sun glint) rather than from other animals. To provide sufficient negative examples, we explicitly keep frames and image regions without dolphins in the training sets, including scenes with surfers or kayakers, fish schools, birds, watercrafts, dolphin looking algae and rocks (Fig. 4). These act as hard negatives and help the detector learn to ignore non-dolphin objects with similar size and contrast. In future Machine Learning with Applications 23 (2026) 100808 4 L. Viegas Filipe et al. Fig. 3. Flowchart illustrating the process of dataset creation and how the annotation tool (Section 5) aids in the development. The top blue box illustrates the creation of the detection dataset using a human-in-the-loop procedure. The bottom green box represents the creation of the tracking dataset. The middle orange box summarizes the capabilities of the annotation tool, linking them to the tasks they aid in the creation of both datasets. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.) Fig. 4. Examples of images with negative samples, e.g., images with no dolphins, surfers or kayakers mixed with dolphins, fish schools, birds, watercrafts, dolphin looking algae and rocks. work, as more multi-species data become available, we plan to extend the dataset and annotations to additional marine species to further stress-test cross-species generalization. Implications for domain shift are discussed in Section 8. 5. Annotation tool To facilitate efficient correction of detections and tracks, we developed a custom annotation tool implemented in PyQt5 python framework. The tool integrates manual labelling, model-assisted refinement, and identity management. The main innovation of the tool lies in the integration of functionalities that were previously scattered across different annotation platforms. We combined the most useful features found in widely used tools such as CVAT (CVAT.ai Corporation, 2024), VIA (Dutta & Zisserman, 2019), and DarkLabel (Programmer, 2025) into a single environment designed to make the annotation process more effective. By consolidating these capabilities into one system, the created tool directly addresses the specific needs of object tracking annotation, significantly reducing manual effort while increasing flexibility in the annotation process. The middle orange box in Fig. 3 illustrates how the tool functionalities directly interact with the annotation processes. Fig. 5 shows the graphical interface of the tool: it consists of two main elements: the frame display and the side menu. The display shows the frame and respective imported annotations. The user annotates boxes and IDs in this area. The side menu consists of buttons with specific functions, e.g., import and export annotations (‘‘Import/Generate Ground-Truth CSV’’, ‘‘Import/Generate predictions CSV’’), and display toggles (‘‘Display predictions/IDs’’). Machine Learning with Applications 23 (2026) 100808 5 L. Viegas Filipe et al. Fig. 5. Graphical interface of the annotation tool. 6. Implementation 6.1. Dolphin detector training and performance evaluation We trained a YOLO11 large (YOLO11L) object detection model to detect dolphins in UAV imagery. A pretrained YOLO11L model was used as the starting point, exploiting transfer learning to accelerate convergence and improve performance compared to training from scratch, and chosen as a balanced compromise between detection accuracy on small objects and computational feasibility. Training was carried out using the Ultralytics YOLO implementation (Jocher et al., 2022) with default hyperparameters, except for the specific adjustments detailed below. The training dataset followed the YOLO format and was specified in a custom YAML file containing image paths and class definitions. Images were resized to 640 × 640 pixels before being passed to the model. Training was performed for 17 epochs on a single NVIDIA GeForce RTX 3060 GPU. To ensure reproducibility, a fixed random seed (42) was used across all runs. Data augmentation was enabled to improve model generalization. Specifically, we applied horizontal flips (fliplr = 0.1), vertical flips (flipud = 0.1), and slight adjustments to hue (hsv_h = 0.1) and brightness (hsv_v = 0.1). These augmentations are particularly relevant in marine environments, where dolphins may appear at varying orientations, under changing lighting conditions, and against heterogeneous sea backgrounds. Flipping augmentations are also important in UAV footage due to the rotational and translational freedom of the camera. In preliminary experiments we found that standard mosaic augmentation decreased detection performance on our UAV dolphin dataset. A likely explanation is that combining frames into a single composite image reduces the effective resolution of already small dolphin targets and introduces sharp seams between sub-images with different sea states, illumination and flight altitudes. This produces unrealistic water textures and partially cropped dolphins or wakes, effectively increasing label noise. Since such artefacts do not appear in deployment imagery, the detector may overfit to mosaic-specific patterns rather than the fine-scale contrast cues that distinguish dolphins from wave crests and whitecaps, which in turn degrades performance on real UAV footage. Therefore, we kept YOLO defaults, avoiding mosaic; simple geometric/photometric augmentations sufficed. Our simplified, ecologically realistic training preserved generalizability. We report mAP@50 and mAP@50:95, reflecting lenient versus Table 2 Main hyperparameters used to train the YOLO11L detector. Hyperparameter Value Input image size 640 × 640 Batch size 32 Number of epochs 17 Optimizer AdamW Initial learning rate 2 ×10−3 (cosine decay) Weight decay 5 ×10−4 Learning rate warm-up 3 epochs Confidence threshold 0.5 NMS IoU threshold 0.3 Data augmentation flips, colour jitter, brightness variation stringent Intersection over Union (IoU) thresholds for localization and ranking (Solawetz, 2025). To further investigate the regions that contribute most to dolphin detection, we generated importance maps using the approach proposed by Petsiuk et al. (2018). To better understand the limitations of the system, we performed a dedicated error analysis of the detector. Hyperparameters. The YOLO11L detector was trained on the detection split of our dataset with the hyperparameters summarized in Table 2. We used standard data augmentation (horizontal flips, colour jitter, random cropping) and selected the final confidence and NMS thresholds on the validation set to balance precision and recall for downstream tracking. 6.2. BoT-SORT optimization and ReID integration 6.2.1. Grid search (coarse exploration) Before applying evolutionary optimization, we conducted an initial grid search to explore the sensitivity of BoT-SORT to key hyperparameters. This was carried out on the AIMM dataset (Section 4.4), chosen to reduce computational cost while still providing indicative results for parameter selection. The parameters explored were: •track_high_thresh in {0.25, 0.5, 0.75}, •track_low_thresh in {0.1, 0.2}, •new_track_thresh in {0.25, 0.5}, •track_buffer in {30, 60} frames, and Machine Learning with Applications 23 (2026) 100808 6 L. Viegas Filipe et al. •match_thresh in {0.8, 0.9}. The other parameters were fixed to default values. All possible combinations of these values were enumerated, created and passed to BoT-SORT. Each configuration was executed on the AIMM dataset, and tracking outputs were stored for later comparison. This procedure allowed us to identify promising regions of the parameter space, which were subsequently injected as seeded individuals in the genetic algorithm optimization. 6.2.2. Genetic algorithm (fine exploration) To refine the results of the grid search, we optimized BoT-SORT hyperparameters using a genetic algorithm (GA) implemented with the DEAP framework. The parameters targeted were the same thresholds and buffers but in a continuous and expanded interval of possible values: •track_high_thresh in (0.2–0.8), •track_low_thresh in (0.1–0.5), •new_track_thresh in (0.2–0.8), •track_buffer in (10–90 frames), and •match_thresh in (0.1–0.8). The GA was initialized with a mixed population of randomly sampled individuals and seeded parameter sets from the grid search. Each individual represented a candidate configuration. The evolutionary process employed blend crossover (𝛼= 0.4), Gaussian mutation (𝜎= 0.1, mutation probability = 0.2), and tournament selection (tournament size = 3). The algorithm ran for 10 generations with a population size of 100 and a Hall of Fame archive of 8 individuals. Each candidate configuration was evaluated by running BoT-SORT on the AIMM dataset (Section 4.4) using YOLO11 detections as input. Performance was assessed with the TrackEval framework (Luiten & Hoffhues, 2020) using the Higher Order Tracking Accuracy (HOTA) metric (Luiten et al., 2020), as well as IDentification F1-score (IDF1). HOTA jointly evaluates detection, association, and localization to provide a single, unified tracker comparison metric. IDF1 evaluates how long the tracker correctly identifies an object and is based on the Identification Precision and Identification Recall. The overall fitness score was defined as: Fitness =0.6⋅HOTA +0.4⋅IDF1.(1) In preliminary experiments, we used only HOTA as the fitness signal for the GA, since it jointly evaluates detection and association quality. However, we observed that some parameter configurations with similar (or slightly higher) HOTA still produced clearly inferior identity preservation in practice, with more fragmented tracks and ID switches. To address this, we extended the fitness to also include IDF1, a widely used identity-centric tracking metric, and combined the two after normalization as presented in Eq. (1). This choice reflects the fact that HOTA remains our primary objective (overall tracking quality), while IDF1 explicitly regularizes the search against solutions that achieve small HOTA gains at the cost of degraded identity consistency. In our experiments, moving from a HOTA-only objective to this two-term fitness systematically reduced severe identity fragmentation with negligible loss in HOTA, and better aligned optimization with our ecological goal of accurately counting unique individuals across the sequence. Hyperparameters. Table 3 summarizes the main BoT-SORT hyperparameters before and after GA optimization. The GA search was run on the tracking validation set using a multi-metric fitness that combines HOTA and IDF1, as described earlier in this section. Unless otherwise stated, all tracking results in the paper use the GA-optimized motion-only configuration. Table 3 Key BoT-SORT hyperparameters. Association and lifespan track_high_thresha– track_low_thresha– new_track_thresha– track_buffer (frames)a– match_thresha– fuse_score True gmc_method sparseOptFlow proximity_threshold 0.5 GA configuration Population size 100 Number of generations 10 Crossover/mutation probabilities 0.4/0.1 Fitness metrics HOTA, IDF1 a Values were obtained from the GA search. Model naming and performance evaluation. Throughout this paper we named the default BoT-SORT coupled with the trained YOLO11 detector as BS. The GA-optimized BoT-SORT coupled with YOLO11 is referred to as BS_GA. All tracking configurations are evaluated on the AIMM dataset and BS as baseline comparison. 6.2.3. ReID integration (appearance cues) To evaluate whether appearance cues could improve dolphin identity tracking, we trained ReID networks using our tracking dataset converted into ReID format. These networks are capable of discriminating visual features and re-identifying individuals based on these cues. This dataset consisted of cropped dolphin detections extracted from corrected tracks; each assigned an identity label. The data were divided into training and validation subsets. Prior to training, we applied a cleaning procedure to improve data quality. This involved (i) removing identities with fewer than 10 usable images, (ii) discarding images that were too small (short side <64 pixels) or excessively blurry (low Laplacian variance), and (iii) eliminating near-duplicate images within each identity using perceptual hashing, keeping only the highest-quality instance. We implemented a threshold-sweeping step to explore different parameter values for minimum image size, blur variance, identity size, and duplicate tolerance. By precomputing image metrics and simulating multiple cleaning configurations, we obtained summary statistics of images and identities retained under each scenario, which allowed us to select thresholds that balanced data quality with dataset coverage. We trained and evaluated several ReID variants within the BoTSORT framework using the deep-person-reid (Torchreid) framework (Zhou & Xiang, 2019). In all cases, detections from the YOLO11L detector were cropped with a small padding margin, resized to the input resolution of the ReID backbone, and normalized using standard ImageNet statistics before being passed through the network to obtain an embedding vector for each detection. We considered two families of backbones commonly used in person re-identification and in existing BoT-SORT implementations, adapted to our single-class marine mammal setting: •OSNet-based models, initialized with publicly available personReID weights. •ResNet50mid, following the mid-level feature variant widely used in tracking-by-detection pipelines. For both backbones, we replaced the final classification layer by a dolphin-specific classifier and fine-tuned the networks on cropped dolphin detections extracted from the training split of our dataset. Training followed a standard protocol combining cross-entropy loss on the identity labels with a batch-hard triplet loss on the embeddings, using random horizontal flips, colour jitter, and mild geometric perturbations as data augmentation. The resulting appearance embeddings Machine Learning with Applications 23 (2026) 100808 7 L. Viegas Filipe et al. Fig. 6. YOLO11 performance on the validation set over the training epochs. Each metric value represents the best result obtained in each epoch, respectively. The notation mAP𝑣𝑎𝑙 𝑡 represents the obtained mAP at a given IoU threshold 𝑡 on the validation set. were then integrated into BoT-SORT, where cosine similarity between embeddings was combined with motion cues during data association. We evaluated these appearance models in two configurations: (i) a full BoT-SORT tracker that combines motion and appearance cues, and (ii) a motion-only baseline in which the appearance term is disabled and associations rely solely on Kalman filter predictions and IoU-based matching. ReID training hyperparameters. Both OSNet and ResNet50mid ReID models were fine-tuned on cropped dolphin detections using a batch size of 32, an initial learning rate of 3×10−4 with single step scheduling with 20 stepsize, and 100 maximum training epochs. We optimized the networks with AdamW, using weight decay of 4×10−4 and a combined loss consisting of cross-entropy on identity labels and batch-hard triplet loss on the embeddings. All ReID experiments reported in this section use this configuration. 7. Results and discussion 7.1. Detection (YOLO11L) The training dynamics reveal a consistent improvement in detection performance across epochs, as summarized in Fig. 6. Validation precision increased quickly, indicating that the model rapidly achieved a high discrimination capacity. Recall followed a steady growth, highlighting enhanced sensitivity to positive detections. The mAP@50 metric exhibited a clear upward trajectory, confirming strong convergence and stable performance. In parallel, the loss functions (Fig. 7) decreased smoothly, with the training losses (Fig. 7(a)) remaining consistently low throughout and validation losses (Fig. 7(b)) stabilizing after the initial variance observed in the classification loss. These results demonstrate that the model reached convergence within the first ten epochs, with subsequent training epochs providing marginal refinements, leading to an overall balance of high precision, recall, and generalization. Visual inspection of a few examples provides insight into how the model is functioning internally. Fig. 8 shows two selected cases that illustrate the generalization capabilities of the detection model. Fig. 8(a) demonstrates the ability of the model to identify small individuals with confidence scores above 0.7. The rightmost detection shows the model correctly detecting the individual, regardless of the glare that affects the area. Fig. 8(b) confirms this robustness, showing detection of dolphins under glare, while also illustrating a good detection where the individual is partially within the frame, further suggesting that the model is resilient to incomplete visual cues. In addition, detections with Table 4 Tracking performance utilizing two distinct BoT-SORT parametrization: default and the resultant from GA optimization. The column IDs represents the total count of dolphins individuals within the video. The annotators agreed on a total of 62 dolphins. The detector was the trained YOLO11L (Section 6.1). BS stands for BoT-SORT. Model HOTA MOTA IDF1 IDs BS (default) 0.631 0.720 0.734 101 BS_GA (ours) 0.636 0.723 0.740 72 lower confidence values (around 0.5–0.6) observed in some examples indicate that the model is still capable of providing useful localization even under suboptimal visibility, though at a reduced certainty. Overall, these examples suggest that the model generalizes well across variations in illumination, scale, and framing, all of which are common in drone-based marine surveys. The resulting importance map visualizations consistently highlighted the dorsal fin, rostrum, and caudal peduncle as the most influential features for the detector’s predictions, regardless of body orientation (Fig. 9). In cases where the animals were partially occluded by water splashes, the model shifted its focus to these more distinct anatomical features, indicating a robustness to noise. Fig. 10 illustrates the complementarity between the importance analysis and the detection output. The importance map (Fig. 10(a)) highlights that the model focuses predominantly on distinctive features of the dolphin body, particularly the dorsal fin and rostrum, which seem to be the most informative cues for identification. At the same time, parts of the background within the detection box are also activated, albeit with much lower weights. This suggests that some residual importance is given to the surrounding water, possibly due to reflections and local contrast patterns. The corresponding detection output (Fig. 10(b)) confirms that this importance is sufficient to support accurate localization, with high confidence score (above 0.9). Taken together, these results indicate that the model reliably detects dolphins while grounding its predictions in meaningful anatomical features, although some sensitivity to background artefacts remains. 7.2. Tracking (GA optimized BoT-SORT without ReID) Table 4 reports the performance of the baseline BS and the proposed BS_GA on the AIMM dataset. The proposed method shows consistent improvements across all metrics. HOTA and Multi-Object Tracking Accuracy (MOTA)1 increased slightly. IDF1 also improved, reflecting better identity preservation. Regarding the total number of distinct IDs, the baseline produced 101 identities compared to the ground-truth 62, while the proposed method reduced this to 72, bringing it much closer to the reference. This reduction in over-fragmentation of tracks is the most significant improvement, indicating enhanced stability and continuity in identity assignment. The evolutionary process showed a rapid improvement in the early generations, with the average fitness rising from 53.7 in generation 1 to nearly 60 by generation 5, as shown in Fig. 11. After this stage, the population converged quickly, with only minor improvements up to generation 10. The maximum fitness remained stable around 60 in the first six generations, only slightly increasing to 60.5 thereafter. This premature convergence can be attributed to the injection of the best individuals found in the gridsearch into the population, which accelerated the rise in average fitness. Although the absolute gains in standard tracking metrics appear modest (e.g. small increases in HOTA and IDF1), the GA search consistently reduced identity fragmentation and ID switches compared with 1MOTA is a summary metric that penalizes false positives, missed targets, and identity switches, aggregating these errors into a single higher-is-better score for multi-object tracking. Machine Learning with Applications 23 (2026) 100808 8 L. Viegas Filipe et al. Fig. 7. Evolution of YOLO11 training and validation losses over the epochs. YOLO11 minimizes three distinct loss functions: box_loss is a regression loss for bounding box position; cls_loss stands for object classification loss; and dfl_loss stands for distribution focal loss and is designed to handle class imbalance and enhance performance on challenging detection scenarios. Fig. 8. Examples of YOLO11 predictions illustrating good performance (a) under glare and (b) with partial dolphin cues. Fig. 9. Two YOLO11 importance maps, generated using an occlusion-based approach, showing how the detector tends to prioritize biologically relevant features. Each map shows the importance only on the most confident detection. the default BoT-SORT configuration, with around a 29% reduction in fragmented tracks in our experiments. This difference is important in our application, where ecological indicators (e.g. abundance estimates, re-sightings) depend on maintaining consistent identities over long sequences rather than on per-frame accuracy alone. From a practical standpoint, the GA optimization is a one-off, offline cost: once a good parameter set is learned for a given sensor and survey protocol, it can be reused across many campaigns with no further tuning. In contrast, manual hyperparameter search is labour-intensive and difficult to reproduce. We therefore consider the additional complexity of GA tuning justified in operational settings where identity preservation and robust, long-term counting are critical, while acknowledging that for quick exploratory studies or applications tolerant to moderate identity noise, the default tracker configuration may be sufficient. 7.3. Tracking (BoT-SORT with ReID) Quantitative results indicated that appearance-assisted BoT-SORT did not improve tracking compared to the default configuration on the AIMM dataset. Metrics remained unchanged or slightly worsened, Machine Learning with Applications 23 (2026) 100808 9