Comparative assessment of YOLO-OBB models for AI-enabled mussel raft detection in VHR remote sensing imagery: insights from the Ría de Arousa, Spain

Abstract

The rapid expansion of mussel raft aquaculture has increased the demand for accurate and scalable monitoring systems to support sustainable coastal management. Deep learning (DL) object detectors, particularly the YOLO family, enable automated extraction of aquaculture infrastructures from Very High-Resolution (VHR) satellite imagery. However, conventional detectors rely on horizontal bounding boxes that often include excessive background when objects are elongated or rotated, reducing localization accuracy. Oriented bounding box (OBB) detectors address this limitation by modeling object orientation and geometry. In this study, an open-source Python-based graphical user interface (GUI) was developed for automated mussel raft detection, and a comprehensive comparison of YOLOv8-OBB and YOLO11-OBB architectures was conducted using VHR imagery from Spain’s Ría de Arousa. A dedicated mussel raft dataset was compiled and annotated, and both model families were evaluated across five architectural scales (n, s, m, l, and x). Experimental results showed that the YOLO11 family generally outperformed YOLOv8 in terms of detection accuracy, localization quality, and performance under challenging maritime conditions. Among the evaluated models, YOLO11s achieved the highest F1-score (0.866) and mAP50–90 (85.21%), while YOLO11n achieved the highest Mean IoU (0.872) and mAP50 (96.87%). The results further demonstrated that compact architectures can provide a favorable balance between accuracy and computational efficiency, as larger models did not consistently yield superior performance. Large-scale experiments on a 60,309 × 61,569-pixel (50 cm resolution) VHR scene demonstrated the operational scalability of the proposed framework through a patch-based processing strategy. The comparative evaluation conducted in this study provides valuable insights into recent YOLO-OBB architectures and highlights their potential for scalable AI-based aquaculture monitoring.

1 Introduction

Mussels, particularly Mytilus galloprovincialis, play a vital role in aquaculture by supporting coastal economies, contributing to food security, and providing ecosystem services as biofilters that enhance water quality and biodiversity (Caballero Miguez et al., 2009; Pérez-Camacho et al., 2013; Rodríguez, 2017; Soliño and Figueras, 2025). These ecological and socioeconomic benefits have driven the global expansion of mussel farming, which is adapted to local environmental conditions through bottom culture, suspended systems (off-bottom; e.g., rafts), and integrated approaches (McKindsey et al., 2011; Mascorda Cabre et al., 2021). Spain ranks third worldwide in mussel production, contributing ∼44.6% of European and ∼10% of global output, with Galicia alone accounting for nearly 90% of national production (Soliño and Figueras, 2025). Within Galicia, the Ría de Arousa is the principal hub, hosting more than of thousands of floating mussel raft cultures (locally known as bateas) and dominating regional production (Caballero Miguez et al., 2009; Pérez-Camacho et al., 2013; Soliño and Figueras, 2025). Effective management of this sector requires precise spatial information for concession licensing, environmental monitoring, hydrodynamic modeling, and maritime safety (Martín-Rodríguez et al., 2024; Amo-Seco et al., 2025).

Conventional monitoring approaches, such as field surveys and boat-based inspections, are costly, labor-intensive, and impractical for regional-scale assessments. Camera-based methods using buoys or shore stations offer localized observations but lack spatial coverage (McMillan et al., 2023; Narang et al., 2024; Zhao et al., 2024). Space/air-borne Remote Sensing (RS), by contrast, offers synoptic and repeated observations, making it more suitable for mapping the distribution of mussel farms across broad coastal areas. For example, Martín-Rodríguez et al., 2024, employed classical Machine Learning (ML) algorithms, including multilayer perceptrons (MLPs), support vector machines (SVMs), and bagged trees (BTs), to detect mussel rafts in Galicia using spectral indices derived from Sentinel-2 imagery. However, the 10–20 m spatial resolution of Sentinel-2 limits the accurate identification and enumeration of individual rafts. This creates a need for approaches capable of detecting and characterizing mussel rafts over wide areas using very high-resolution (VHR) imagery. In addition, recent advances toward intelligent ocean frameworks highlight the growing role of AI in integrating multi-source marine observations and enabling scalable, automated monitoring of coastal systems. Within this context, VHR imagery provides the spatial detail required for reliable mapping of aquaculture infrastructures and forms a critical component of data-driven, AI-enabled marine observation pipelines. Unlike traditional ML approaches that rely on handcrafted features, Deep Learning (DL) models automatically learn hierarchical feature representations (Moghimi et al., 2024), making them well suited to heterogeneous and dynamic aquaculture settings. Among DL techniques, the YOLO (You Only Look Once) family has gained prominence for balancing computational efficiency and detection accuracy (Redmon et al., 2016). YOLO-based models have been successfully applied in various marine science domains, including microplastic detection and classification (Akkajit et al., 2024), ship detection (Ricci and Bohnenstiehl, 2022; Bakirci, 2025), corrosion detection (Zhou et al., 2022) and oil spill monitoring (Cai et al., 2024). The strength of YOLO lies in its single-pass architecture, which enables rapid processing of large-area imagery without sacrificing accuracy. This capability is particularly relevant for aquaculture monitoring, as it facilitates the detection of the regular geometric patterns characteristic of floating mussel rafts (Redmon et al., 2016; Narang et al., 2024). This potential also illustrated in an Esri Developer tutorial using the ArcGIS API for Python (Esri, 2025). However, standard YOLO variants [e.g., YOLOv8 (Yaseen, 2024; Varghese and Sambath, 2024) and YOLOv11 (Khanam and Hussain, 2024)] generate axis-aligned bounding boxes (AABB), which often fail to capture the geometry of elongated or oblique mussel rafts (see Figure 1b). Such misrepresentations can lead to inaccurate centroid locations and area estimates, which may negatively affect concession compliance, environmental assessments, and maritime navigation. Oriented bounding box (OBB) extensions of YOLO, recently developed by Ultralytics¹ (Ding et al., 2019; Jocher, 2025) address this limitation by predicting rotated bounding boxes that align more closely with raft geometry (see Figure 1a). To the best of our knowledge, these OBB-based YOLO models have not yet been applied to mussel raft detection.

FIGURE 1

To bridge this gap, we developed an open-source graphical user interface (GUI) software (Software Link) that integrates YOLOv8-OBB and YOLOv11-OBB for detecting mussel rafts in VHR imagery accessible through platforms such as SASPlanet². The software allows selection among model sizes [nano (), small (), medium (), large (), extra-large ()], enabling trade-offs between computational efficiency and detection accuracy. Its GUI design reduces technical barriers, making advanced DL accessible to aquaculture managers, researchers, and policymakers. Furthermore, we provide fine-tuned weights and a curated dataset of over 403 annotated VHR image patches of mussel rafts from Galicia, all publicly available for reuse and benchmarking (Dataset Link). By systematically comparing YOLOv8-OBB and YOLOv11-OBB over multiple weight configurations, our framework identifies the most efficient architectures for mussel raft detection and establishes a transferable, open-source solution for mussel rafts monitoring.

2 Materials and methods

2.1 Study area

The present study was conducted in the Ría de Arousa, one of the Rías Baixas located in Galicia, northwestern Spain (see Figure 2). The estuary covers approximately 230 km² and is widely recognized as one of Europe’s most productive coastal environments due to favorable hydrographic conditions like high nutrient input, tidal mixing, and sheltered morphology (Martínez et al., 2011). The region hosts the highest concentration of mussel rafts in Galicia, representing nearly 70% of the platforms distributed across the Rías Baixas (Martínez et al., 2011; NASA Earth Observatory, 2022).

FIGURE 2

2.2 Dataset

VHR satellite imagery was acquired using the SASPlanet environment from publicly available Bing Maps imagery. The imagery was collected at zoom levels 19–20, corresponding to an approximate ground sampling distance (GSD) of 0.15–0.50 m, which allowed individual mussel rafts to be visually identified and delineated. The imagery was subsequently divided into nonoverlap 640 × 640 pixel patches to support DL–based object detection while maintaining sufficient spatial context around raft structures. The original dataset consisted of 403 manually annotated VHR image patches collected from the Ría de Arousa, Spain. The dataset included both raft and non-raft regions. Non-raft samples contained coastal background features such as shoreline structures, vessels, sea clutter, and rocky coastal textures. The dataset was divided into 174 training patches (129 raft/45 non-raft), 49 validation patches (39 raft/10 non-raft), and 180 independent test patches (159 raft/21 non-raft). Data augmentation was applied only to the training subset and included combinations of horizontal flipping, vertical flipping, and rotational transformations (±10° and ±90°). Two augmented variants were generated from each original training patch, increasing the number of training samples from 174 to 522 patches. No augmentation was applied to the validation or independent test subsets in order to keep the evaluation procedure independent from the training data. The larger independent test set was used to provide a more reliable evaluation across different coastal and environmental conditions within the study area.

2.3 Methodology

This study presents a systematic framework for mussel raft detection from VHR satellite imagery using OBB-based YOLO detectors. The overall workflow, including data preparation, annotation, model training, inference, evaluation, and deployment through an open-source Python-based graphical user interface (GUI), is illustrated in Figure 3. The dataset described in Section 2.2 was used for model development and independent evaluation.

FIGURE 3

Let the dataset be , where denotes the total number of annotated image patches, represents th RGB image patch, and denotes corresponding set of ground-truth OBBs, with indicating the number of mussel rafts in image . Each oriented bounding box is parameterized as:where represent the center coordinates, and denote the width and height, respectively, and is the orientation angle. The dataset was partitioned into training, validation, and testing subsets:

Such that:

Mussel raft detection was performed using OBB-enabled variants of the YOLO family, specifically YOLOv8 (Yaseen, 2024; Varghese and Sambath, 2024) and YOLOv11 (Khanam and Hussain, 2024) implemented through the Ultralytics³ framework (Jocher, 2025) (see Section 2.3.1; Section 2.3.2). Each model family provides multiple scale configurations , corresponding to increasing model capacity and computational complexity. Let denote a detector of type , and scale , parameterized by learnable weights . For a given input image the detector produces a set of predicted bounding boxes:where , and each predicted box follows the same parameterization as the ground truth.

The detector was trained by minimizing a composite loss function, denoted as that rotated bounding box regression (), classification (), and distribution focal terms (), experesed bywhere are weighting coefficients. The that measures geometric discrepancy using rotated Intersection over Union (), defined as:where measures the geometric overlap between predicted and ground-truth oriented bounding boxes.

Given the single-class nature of the problem (mussel raft), the is formulated as binary cross-entropy:where indicates the presence of a mussel raft and is the predicted probability.

To improve localization accuracy, particularly along object boundaries, is employed.

For a target distance , let , , and be the predicted probability for bin , is then defined as:

Training is conducted on , while was used to monitor validation performance and select the optimal model checkpoint following the procedure described in Algorithm 1. The final evaluation was conducted on the independent test set using metrics mentioned in Section 2.4. The evaluation procedure follows the workflow outlined in Algorithm 2.

To support reproducibility and practical usability, the proposed framework was implemented as an open-source Python-based GUI application. The interface allows users to load satellite imagery, perform inference using trained YOLO-OBB models, and visualize predicted oriented bounding boxes.

2.3.1 YOLOv8-OBB

The YOLOv8-OBB model employed in this study follows the standard detection architecture proposed by Ultralytics, extending the original YOLOv8 framework with an oriented bounding box (OBB) formulation for rotation-aware object detection (Jocher, 2025). As illustrated in Figure 4, the architecture comprises three main components: a backbone for feature extraction, a neck for multi-scale feature fusion, and a detection head for OBB prediction.

FIGURE 4

Jun et al., 2025).

The backbone consists of convolutional layers integrated with Cross Stage Partial (CSP)-inspired C2f modules, which promote efficient feature reuse and stable gradient propagation. This design supports robust feature representation while maintaining computational efficiency. At the final stage, a Spatial Pyramid Pooling Fast (SPPF) module aggregates contextual information across multiple receptive fields, enhancing the representation of objects at different scales (Jun et al., 2025). The neck adopts a feature pyramid network (FPN)-like structure to fuse spatial and semantic features across scales, improving robustness to variations in object size and orientation. The detection head operates on multiple feature levels and predicts class probabilities together with oriented bounding box parameters .

2.3.2 YOLOv11-OBB

YOLOv11-OBB represents a more recent iteration of the YOLO detection framework, designed to further enhance feature representation for rotation-aware object detection (Jocher, 2025). While preserving the backbone–neck–head paradigm, it introduces architectural refinements that improve contextual modeling and feature discrimination, as illustrated in Figure 5.

FIGURE 5

Xu et al., 2025).

The backbone builds upon convolutional layers with C3K2 modules, extending the CSP design to improve feature interaction and effective depth without significant computational overhead. Similar to YOLOv8-OBB, an SPPF module is used for multi-scale contextual aggregation. In addition, a C2PSA module is incorporated after the SPPF stage to perform attention-based feature recalibration. This self-attention mechanism enhances informative spatial and channel-wise features, enabling improved modeling of long-range dependencies, which is particularly beneficial for detecting elongated and arbitrarily oriented objects in remote sensing imagery (Xu et al., 2025). The neck employs an FPN-like multi-scale fusion strategy to integrate features across resolutions, improving robustness to scale and orientation variations. The detection head follows the same OBB formulation, predicting class probabilities and bounding box parameters at multiple feature scales.

2.4 Evaluation metrics

The model’s performance was assessed using rotated Intersection over Union (rIoU), based matching after rotated non-maximum suppression (NMS). A detection was considered a true positive () if it correctly identified the class (mussel raft) and achieved an with an unmatched ground-truth box (one-to-one matching). Detections without a valid match were counted as false positives (), while unmatched ground-truth objects were counted as false negatives () (Yu et al., 2025). Duplicate detections associated with the same ground-truth object were also treated as . Given the highly oriented structure of mussel rafts, localization quality was assessed using IoU metric, defined as:where, and are the predicted bounding and corresponding ground-truth OBB, respectively.

To mitigate potential bias arising from the uneven spatial distribution of objects across image patches, all global metrics were computed using a micro-averaging strategy. Under this framework, TP, FP, and FN counts were aggregated across the entire test dataset prior to metric calculation. Global Precision and Recall were computed as:

The Global F1-score, representing the harmonic mean of Precision and Recall, was calculated as:

To further evaluate localization accuracy, the Global Mean IoU was computed as the TP-weighted average IoU across all correctly detected objects:

To further assess localization quality, the Precision–Recall (PR) curve was obtained by sweeping the confidence threshold at a fixed rIoU value. The Average Precision (AP) is defined as the area under this curve:

Following standard object detection protocols, mAP@50 was reported at an IoU with , whereas mAP@50:95, was calculated as the mean AP over thresholds ranging from 0.50 to 0.95 with a step size of 0.05:

Since the task involves a single class, mAP reduces to the AP of the mussel-raft class under the specified thresholds.

Model calibration and operating-point sensitivity were additionally assessed using Precision–Confidence, Recall–Confidence, and F1–Confidence curves. All threshold selection and sensitivity analyses were performed exclusively on the validation dataset to prevent information leakage. To quantify uncertainty and evaluate model robustness, bootstrap resampling was applied to the independent test dataset. Specifically, 1,000 bootstrap samples were generated through random sampling with replacement at the image-patch level. For each iteration, Precision, Recall, F1-score, and Mean IoU were recomputed. The corresponding 95% confidence intervals (CIs) were obtained from the 2.5th and 97.fifth percentiles of the bootstrap distributions. Statistical significance between competing methods was also assessed using the Wilcoxon signed-rank test at a significance level of ( = 0.05).

3 Experimental results and discussion

3.1 Experimental setup

All experiments were conducted in Google Colab using an NVIDIA T4 GPU with 14 GB memory. The official GitHub repositories of YOLOv8-OBB and YOLOv11-OBB were cloned into Google Drive, and both models were trained and evaluated within this environment. Implementation was carried out using the Ultralytics YOLO vision liberary⁴, which supports oriented bounding box detection. As discussed before, the dataset comprised 522 training patches, 49 validation patches, and 180 test patches, including both mussel raft and non-raft samples to ensure class balance. Training was performed for 50 epochs using the stochastic gradient descent framework with the AdamW optimizer, configured with a base learning rate of 0.002 and a momentum factor of 0.9. Table 1 summarizes the YOLO-OBB architectures employed in this study.

3.1.1 Performance comparison of YOLO-OBB models during training

The box regression loss (Box Loss) was tracked over 50 epochs on the training and validation datasets to evaluate convergence and generalization. The per-epoch Box Loss curves for all YOLO-OBB models in mussel raft detection are shown in Figure 6.

FIGURE 6

Across all model scales, the training loss exhibited a steep decline during the early epochs, indicating rapid learning and effective localization error minimization, followed by a gradual plateau as bounding box predictions stabilize. The validation loss showed a similar trend, with fluctuations influenced by model size and complexity. Smaller variants (YOLOv8n, YOLOv11n) converged quickly and maintain stable validation losses, making them well-suited for resource-constrained settings. In contrast, Larger variants (YOLOv8x, YOLOv11x) showed higher and more variable validation losses, suggesting a risk of overfitting without sufficient regularization or data diversity. Intermediate models (YOLOv8m, YOLOv11 m) displayed smoother convergence, whereas YOLOv8 and YOLOv11 l balanced convergence with moderate train–validation gaps, reflecting strong generalization.

To further assess detection behavior, four evaluation curves, including Precision–Confidence, Recall–Confidence, F1–Confidence, and Precision–Recall, were generated for all trained models. Figures 7, 8 present the results for the YOLOv8-OBB and YOLOv11-OBB families across the five weight configurations (n, s, m, l, x). For YOLOv8-OBB (Figure 7), the Precision–Confidence curves showed a consistent upward trend, with all variants surpassing 95% precision beyond a confidence threshold of 0.6. YOLOv8s and YOLOv8x exceeded 98% precision near 0.8, while YOLOv8m and YOLOv8L reached slightly lower values due to more false positives. Recall peaked at lower thresholds and declined with increasing confidence, though YOLOv8x and YOLOv8L maintained recall above 95% until ∼0.7, whereas YOLOv8n and YOLOv8m dropped below 80% near 0.9. The F1–Confidence curves indicated optimal performance between 0.4 and 0.6, with YOLOv8s achieving an F1 close to 1.0. Smaller variants (YOLOv8n, YOLOv8m) show slightly reduced peaks, indicating modest trade-offs between precision and recall. Precision–Recall curves confirmed strong detection capacity across all variants, with YOLOv8s and YOLOv8x showing the most stable performance, while YOLOv8n and YOLOv8m displayed greater trade-offs under stricter conditions.

FIGURE 7

FIGURE 8

For YOLOv11-OBB (Figure 8), the Precision–Confidence curves remained consistently high, with YOLOv11s and YOLOv11n maintaining precision above 0.98 up to a threshold of ∼0.85, highlighting their effectiveness in reducing false positives. Larger models (YOLOv11L, YOLOv11x) sustained recall above 95% until ∼0.92, whereas YOLOv11 m declined more steeply, indicating stricter filtering. The F1–Confidence curves showed optimal performance between 0.8 and 0.9, with YOLOv11s and YOLOv11 l peaking near 0.96, while YOLOv11 m and YOLOv11x achieved slightly lower peaks with sharper declines. Precision–Recall curves confirmed overall robustness, with YOLOv11s and YOLOv11 l showing the most stable precision–recall balance, and YOLOv11x displaying earlier curvature, reflecting a modest increase in false positives at high recall.

3.1.2 Threshold calibration and sensitivity analysis

To ensure consistent evaluation across architectures, a fixed confidence threshold of 0.3 was applied in all experiments. This threshold was chosen based solely on the validation set to prevent information leakage and ensure independence in the final test evaluation. Figures 7, 8 show that a confidence threshold of 0.3 balances precision and recall effectively across the YOLOv8-OBB and YOLO11-OBB model families, while maintaining stable F1-scores, indicating low sensitivity to minor variations.

Following confidence calibration, sensitivity analysis was conducted with IoU thresholds of 0.5, 0.6, 0.7, and 0.8, independent of the confidence threshold to avoid bias in the final evaluation protocol. Figure 9 shows that IoU thresholds of 0.5 and 0.6 yield high detection scores (Global F1-score ≥0.90) but do not challenge the models’ ability to extract precise geometric features, reducing their ability to differentiate between basic and advanced models. On the other hand, an IoU threshold of 0.8 considerably reduces the F1-score due to heightened sensitivity to minor localization errors.

FIGURE 9

To further investigate this behavior, Figure 10 presents qualitative examples resulted from using YOLOv8-s from challenging image patches. These examples illustrate that detections with high visual agreement with reference annotations may be rejected under the strict IoU = 0.8 criterion, despite having significant spatial overlap. Conversely, at lower thresholds, these bounding boxes achieve a satisfactory mean overlap (mIoU ≈ 0.82) and are correctly classified as True Positives. This highlights the potential drawbacks of overly restrictive overlap thresholds in evaluating oriented object detection performance.

FIGURE 10

Based on the analyses in Figures 9, 10, an IoU threshold of 0.7 was selected for the final evaluation, providing a good balance between localization accuracy and detection robustness. Accordingly, all results in Section 3.2 use a confidence threshold of 0.3 and an IoU threshold of 0.7 on the independent test dataset.

3.2 Comparison results in mussel raft detection

Table 2 summarizes the detection performance of the YOLOv8-OBB and YOLO11-OBB model families across five architectural scales (n, s, m, l, and x) on the independent test dataset.

The results show differences in performance across both model generations and network scales. In general, the YOLO11 variants achieved higher global metrics, Mean IoU, and mAP50–90 values than most YOLOv8 models, although this trend was not consistent across all scales.

Within the YOLOv8 family, performance improved from the nano to medium scales, with YOLOv8m achieving the highest F1-score (0.776) and mAP50–90 (79.30%). However, further increases in model size did not result in additional gains. For example, although YOLOv8x achieved a relatively high Recall (0.801), its F1-score (0.759) and mAP50–90 (74.24%) remained lower than those of YOLOv8m.

Among the YOLO11 variants, YOLO11n and YOLO11s achieved the strongest detection performance. YOLO11s obtained the highest F1-score (0.866) and mAP50–90 (85.21%), representing improvements of approximately 11.6% and 7.5%, respectively, compared with the best-performing YOLOv8 model (YOLOv8m). Similarly, YOLO11n achieved an F1-score of 0.864 and mAP50–90 of 84.57%, indicating that both compact YOLO11 models outperformed the larger YOLOv8 variants. In contrast, YOLO11 m produced noticeably lower Precision (0.694), F1-score (0.739), and mAP50–90 (74.22%) values than YOLO11n and YOLO11s.

The localization metrics followed a similar pattern. Mean IoU values ranged from 0.798 to 0.869 within the YOLOv8 family and from 0.850 to 0.872 within the YOLO11 family. The highest Mean IoU was obtained by YOLO11n (0.872), followed closely by YOLO11s and YOLO11L, both achieving values of approximately 0.871. Similarly, the highest mAP50 value was achieved by YOLO11n (96.87%), while YOLO11s and YOLO11 l produced comparable values of 96.01% and 96.21%, respectively.

A different trend was observed for YOLO11x. Although it maintained relatively high Precision (0.794) and Recall (0.858), its mAP50–90 decreased substantially to 46.26%, suggesting reduced localization consistency under stricter overlap requirements.

Overall, YOLO11n and YOLO11s achieved the strongest balance between detection accuracy and localization quality among the evaluated models, with YOLO11s providing the highest detection performance and YOLO11n achieving the most accurate object localization.

To further examine model performance, Figure 11 presents the global confusion matrices of the YOLOv8-OBB and YOLO11-OBB models, providing a detailed view of the distribution of TP, FP, and FN over different scales.

FIGURE 11

Within the YOLOv8 family, the number of correctly detected mussel rafts increased from 256 for YOLOv8n to 316 for YOLOv8m, while the number of missed targets decreased from 131 to 71. This trend is consistent with the improvements observed in the global evaluation metrics reported in Table 2. However, the performance gains were not maintained for the larger variants. YOLOv8x produced 310 TP and 77 FN, resulting in slightly lower detection performance than YOLOv8m. In addition, all YOLOv8 models generated more than 100 false positives, indicating a relatively high level of confusion between mussel rafts and surrounding marine background features.

A similar trend was observed for false positives. YOLO11s generated 65 FP, compared with 135 FP for YOLO11m and 105 FP for YOLO11 l. These results indicate that the increase in model scale within the YOLO11 family was not consistently associated with improvements in detection performance. Overall, the confusion matrix analysis is consistent with the quantitative metrics presented in Table 2, where YOLO11n and YOLO11s achieved the highest overall detection performance among the evaluated models.

To produce complementary perspective to the global metrics the patch-level distributions of F1-score and Mean IoU for the evaluated YOLOv8-OBB and YOLO11-OBB models presented into violin plots in Figure 12.

FIGURE 12

In Figure 12a, the F1-score distributions for YOLOv8 variants reveal wider spreads with longer lower tails, indicating greater variability in mussel raft detection performance. YOLOv8m stands out within this group, achieving the highest median F1-score and more concentrated distribution, aligning with its superior metrics in Table 2. In contrast, YOLO11 models show distributions shifted toward higher F1-scores with reduced variability. Among these, YOLO11n and YOLO11s have the highest medians and most concentrated distributions. A similar pattern is observed for the Mean IoU distributions presented in Figure 12b. YOLOv8 models exhibit more variability in localization performance, particularly in the n, s, and l variants, while YOLOv8m has the highest median Mean IoU with a compact distribution. In comparison, YOLO11 models achieved higher median Mean IoU values with narrower distributions. Although YOLO11n reached the highest global Mean IoU of 0.872, YOLO11s maintains one of the most compact distributions, indicating consistent localization performance across image patches.

To further evaluate whether the observed performance differences were statistically significant, pairwise Wilcoxon signed-rank tests were performed on the patch-level F1-score distributions. The Holm–Bonferroni procedure was applied to account for multiple comparisons, and statistical significance was assessed at α = 0.05.

The results presented in Table 3 indicate that most comparisons between YOLOv8 and YOLO11 variants were statistically significant (p < 0.05), confirming substantial differences in patch-level F1-score distributions between the two model generations. In particular, YOLO11n and YOLO11s exhibited statistically significant improvements relative to all YOLOv8 variants. In contrast, several comparisons among models within the same generation were not statistically significant, indicating that differences in model scale did not always translate into meaningful differences in classification performance. These findings are consistent with the quantitative results reported in Table 2, where YOLO11n and YOLO11s achieved the highest F1-scores among the evaluated models.

Table 4 presents the Wilcoxon signed-rank test results for the patch-level Mean IoU distributions. Although several comparisons between YOLOv8 and YOLO11 variants remained statistically significant, a larger number of non-significant comparisons were observed than for F1-score. For example, YOLO11n (p = 0.519), YOLO11s (p = 1.000), and YOLO11 l (p = 1.000) did not exhibit statistically significant differences from YOLOv8m in terms of Mean IoU distribution. Similarly, YOLO11 m and YOLOv8x showed statistically equivalent localization performance (p = 1.000).

Overall, the Wilcoxon test results indicate that the superiority of YOLO11 over YOLOv8 is more evident in detection performance than in localization accuracy. Since several top-performing models exhibited statistically comparable Mean IoU distributions, a qualitative evaluation was also conducted to further examine differences in object delineation, boundary alignment, and orientation estimation. Representative examples are shown in Figures 13, 14 for image patches in which mussel rafts are clearly distinguishable from the surrounding marine background.

FIGURE 13

FIGURE 14

As illustrated in Figures 13, 14, all evaluated models were generally able to detect the presence of mussel rafts. However, differences can be observed in the geometric alignment and orientation of the predicted bounding boxes. Within the YOLOv8 family (Figure 13), the smaller variants, particularly YOLOv8n and YOLOv8s, occasionally produced bounding boxes that included portions of the surrounding water background or exhibited minor deviations from the reference annotations. In comparison, YOLOv8m, YOLOv8L, and YOLOv8x generally produced tighter object boundaries.

The YOLO11 family (Figure 14) produced visually more consistent detections across all model scales. In particular, YOLO11n and YOLO11s generated bounding boxes that closely followed the shape and orientation of the mussel rafts while maintaining limited background inclusion. Similar performances can be observed in scenes containing multiple adjacent or densely distributed rafts, where the predicted orientations remained closely aligned with the dominant axis of the target structures. Although the visual differences between the top-performing models are relatively small, the qualitative results are consistent with the quantitative and distributional analyses presented in the previous sections.

To further investigate model behaviour under challenging environmental conditions, Figures 15, 16 present representative test patches containing complex background features and potential sources of semantic confusion. The selected examples include densely distributed mussel rafts, nearby vessels, shoreline environments, rocky coastal regions, and scenes without any target rafts.

FIGURE 15

FIGURE 16

For the negative examples shown in Figures 15, 16(g–i), where no mussel rafts are present, most YOLOv8-OBB and YOLO11-OBB variants avoided false detections, indicating that rocky coastlines, shallow-water textures, and complex nearshore backgrounds were generally not confused with raft structures. However, the shoreline case in Figures 15, 16e was more challenging. In this case, several YOLOv8 variants produced false detections along coastal or land–water boundary features, with this behaviour being more visible in YOLOv8m, YOLOv8L, and YOLOv8x. In comparison, the YOLO11 family showed fewer shoreline-related false detections, although occasional incorrect predictions in terms of shape and orientation were still observed in most scales except particularly for YOLO11n.

The example shown in Figures 15, 16f represents another source of semantic ambiguity because boats and harbor structures can have elongated shapes similar to mussel rafts. In the YOLOv8 family, false detections were mainly observed in the smaller variants, especially YOLOv8s, where vessel-related structures were incorrectly detected as mussel rafts. In the YOLO11 family, most variants suppressed this type of error, although YOLO11s still produced an isolated false detection near the vessel structure.

For densely distributed mussel rafts, particularly in Figures 15, 16b,d, accurate orientation estimation is important for separating adjacent objects and reducing overlap between predicted bounding boxes. Both model families detected most raft structures in these examples. However, the YOLO11 variants generally produced more consistent oriented bounding boxes, with better alignment to the dominant axis of the raft structures. This behaviour was especially visible for YOLO11n and YOLO11s, which maintained compact boundaries while limiting overlap with neighbouring targets.

In the more isolated raft examples shown in Figures 15, 16a,c, the target objects are visually more distinct from the background. Under these conditions, most models produced acceptable detections. Nevertheless, YOLOv8n and YOLOv8s showed minor boundary expansion or orientation deviations in some cases, whereas YOLO11n and YOLO11s produced tighter and more stable object boundaries.

To further evaluate model behaviour under challenging sea-surface conditions, Figures 17, 18 present representative examples containing wave-induced textures and wake patterns surrounding the mussel rafts. These structures introduce additional background variability and may partially resemble the elongated geometry of the target objects.

FIGURE 17

FIGURE 18

Across most examples, both YOLOv8-OBB and YOLO11-OBB models successfully detected the majority of mussel rafts. However, the presence of strong wave patterns occasionally resulted in false positives, missed detections, or minor boundary deviations. Within the YOLOv8 family, these effects were more frequently observed in the small, medium, and large variants, where several examples exhibited either false detections or incomplete identification of target rafts.

The YOLO11 family generally maintained more consistent detections under the same conditions. In particular, YOLO11n and YOLO11s produced fewer visible false detections and preserved closer alignment between the predicted oriented bounding boxes and the reference annotations. Nevertheless, some localization inaccuracies and missed detections remained visible in a limited number of examples, indicating that wave-induced background patterns continue to represent a challenging scenario for both model generations.

4 Discussion and practical implications

The experimental results revealed several important observations regarding the applicability of YOLO-based OBB architectures for mussel raft detection in VHR imagery. To facilitate interpretation of the overall findings, Table 5 presents an integrated comparison of the evaluated models across accuracy, localization quality, computational efficiency, and performance under challenging maritime conditions.

As can be seen form Table 5, most of YOLO11 variants generally achieved better detection accuracy and more consistent localization behavior than the YOLOv8 models. Similar findings have been documented in recent studies, where YOLO11-based detectors demonstrated enhanced efficiency and accuracy over previous YOLO generations in RS object detection (Shi et al., 2024; Zhu and Ablameyko, 2026). These gains may be attributable to architectural enhancements such as the C3k2 block and C2PSA module, which improve feature representation and contextual information extraction (Hidayatullah et al., 2025). For mussel raft detection, where targets are elongated, densely distributed, and often surrounded by visually complex marine backgrounds, improved feature discrimination and localization consistency are particularly important.

Additionally, our results indicate that increasing model size does not necessarily improve mussel raft detection performance. Larger architectures, despite having more parameters and computational complexity, did not consistently outperform the compact variants. Within the YOLO11 family, YOLO11n and YOLO11s matched or outperformed several larger models. Qualitative analysis showed that larger models occasionally exhibited over-segmentation in dense raft environments, leading to redundant bounding boxes or orientation inconsistencies. This aligns with other object-detection studies suggesting that beyond a certain point, increasing model capacity offers limited advantage when target objects have regular geometric patterns (Hu et al., 2023; Sangha and Darr, 2025; Yang et al., 2025). Given that mussel rafts possess repetitive and well-defined structures, compact models may suffice to capture the necessary discriminative information, meaning that larger models could introduce unnecessary complexity without proportional performance gains.

In challenging cases, detection errors predominantly occurred in shoreline areas, wave-affected textures, and vessel-related structures, which share elongated shapes and repetitive patterns or radiometric similarities with mussel rafts, increasing false detection and localization errors. Such confusion is typical in maritime object detection and aquaculture monitoring systems. Conversely, rocky coastal regions and simpler environments resulted in fewer errors for both model types. These findings underscore the necessity of evaluating aquaculture monitoring systems under varied conditions rather than relying solely on aggregate performance metrics. Overall, the integrated assessment in Table 5 highlights that while several architectures excelled in specific areas, only a few consistently achieved high ratings across all criteria. The compact YOLO11n and YOLO11s variants provided one of the most balanced trade-offs between accuracy, localization quality, efficiency, indicating that increased model complexity is not always advantageous for practical performance. Based on our experiments and datasets, Based on the experimental results, YOLO11n appears particularly suitable for large-scale monitoring or resource-constrained environments due to its computational efficiency, whereas YOLO11s provides a slight advantage in detection accuracy and localization quality.

An additional practical aspect concerns the applicability of the proposed framework to large-area monitoring. To investigate operational scalability, the developed software was applied to a VHR image of Galicia’s Ría de Arousa covering approximately 60,309 × 61,569 pixels at 50 cm spatial resolution (Figure 19). Both YOLOv8x-OBB and YOLO11 l-OBB were successfully deployed through a patch-based processing strategy, demonstrating the feasibility of large-scale mussel raft mapping. Nevertheless, some false alarms remained visible, primarily associated with coastal structures, vessels, floating objects, and wave-induced sea-surface patterns that partially resembled raft geometries. Among the two evaluated models, YOLO11 l-OBB produced fewer false detections than YOLOv8x-OBB, suggesting more consistent behaviour in the analysed large-scale scene. These results indicate that the proposed framework is not limited to small image patches and can potentially support operational mussel raft inventories over extensive coastal regions. However, further evaluation across different acquisition dates, environmental conditions, and geographic regions is still required to assess long-term applicability and geographic transferability.

FIGURE 19

A subsequent qualitative evaluation examined model applicability beyond the training region using YOLO11s-OBB on a Southeast Asian aquaculture area (see Figure 20). As can be seen form Figure 20 Although the model successfully detected several raft structures with geometric characteristics similar to those observed in Galicia, a number of structures with substantially different shapes, sizes, and spatial arrangements were missed. These observations suggest that additional fine-tuning or the inclusion of geographically diverse training samples may improve adaptability when transferring the framework to regions with different aquaculture layouts.

FIGURE 20

5 Conclusion

This study presented a comprehensive evaluation of YOLOv8-OBB and YOLO11-OBB architectures for mussel raft detection in VHR imagery. A dedicated mussel raft dataset and an open-source Python-based GUI software were developed to support practical implementation and large-scale mapping. The findings demonstrated that OBB-based detectors provide an effective framework for extracting elongated aquaculture structures from VHR imagery while maintaining a balance between detection performance and computational efficiency. Among the evaluated architectures, most YOLO11 variants achieved stronger overall performance than their YOLOv8 counterparts in terms of quantitative accuracy, localization quality, and robustness under challenging conditions. In particular, compact YOLO11 variants like version n and s provided a favourable trade-off between detection performance and computational efficiency, indicating that increasing model complexity does not necessarily improve practical applicability. Large-scale experiments in Galicia’s Ría de Arousa demonstrated the operational scalability of the proposed framework and its applicability to extensive VHR imagery through a patch-based processing strategy. The results further suggest that OBB-based object detectors represent a promising solution for automated mussel raft inventory generation and coastal aquaculture monitoring.

Future work should focus on expanding the training dataset with mussel raft samples from different geographic regions and aquaculture systems, investigating cross-region transferability, and evaluating the framework under a broader range of environmental and acquisition conditions.

References

• 1

  AkkajitP.AlahiM. E. E.SukkueaA. (2024). Enhanced detection and classification of microplastics in marine environments using deep learning. Reg. Stud. Mar. Sci.80, 103880. 10.1016/j.rsma.2024.103880

  • CrossRef
  • Google Scholar
  • View reference in article

• 2

  Amo-SecoM.CastroC. G.CollinsC. A.BergP. (2025). Tidal modulation of benthic oxygen fluxes in a suspended mussel aquaculture. Mar. Environ. Res.205, 107003. 10.1016/j.marenvres.2025.107003

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 3

  BakirciM. (2025). Advanced ship detection and ocean monitoring with satellite imagery and deep learning for marine science applications. Reg. Stud. Mar. Sci.81, 103975. 10.1016/j.rsma.2024.103975

  • CrossRef
  • Google Scholar

• 4

  Caballero MiguezG.Garza GilM. D.Varela LafuenteM. M. (2009). The institutional foundations of economic performance of mussel production: the Spanish case of the Galician floating raft culture. Mar. Policy33, 288–296. 10.1016/j.marpol.2008.07.008

  • CrossRef
  • Google Scholar

• 5

  CaiY.ChenL.ZhuangX.ZhangB. (2024). Automated marine oil spill detection algorithm based on single-image generative adversarial network and YOLO-v8 under small samples. Mar. Pollut. Bull.203, 116475. 10.1016/j.marpolbul.2024.116475

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 6

  DingJ.XueN.LongY.XiaG.-S.LuQ. (2019). “Learning RoI transformer for oriented object detection in aerial images,” in 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (Long Beach, CA, USA: IEEE), 2844–2853. 10.1109/CVPR.2019.00296

  • CrossRef
  • Google Scholar
  • View reference in article

• 7

  Esri (2025). Detecting mussel farms using deep learning | ArcGIS API for python | esri developer. Available online at: https://developers.arcgis.com/python/samples/detecting-mussel-farms-using-deep-learning/ (Accessed September 2, 2025).

  • Google Scholar
  • View reference in article

• 8

  JocherG. (2025). Ultralytics YOLO-OBB. Available online at: https://github.com/ultralytics/ultralytics

  • Google Scholar
  • View reference in article

• 9

  JunE. S.SimH. J.MoonS. J. (2025). Advancing YOLOv8-Based wafer notch-angle detection using oriented bounding boxes, hyperparameter tuning, architecture refinement, and transfer learning. Appl. Sci.15, 11507. 10.3390/app152111507

  • CrossRef
  • Google Scholar
  • View reference in article

• 10

  HidayatullahP.SyakraniN.SholahuddinM. R.GelarT.TubagusR. (2025). YOLOv8 to YOLO11: a comprehensive architecture in-depth comparative review. 10.48550/ARXIV.2501.13400

  • CrossRef
  • Google Scholar

• 11

  HuM.LiZ.YuJ.WanX.TanH.LinZ. (2023). Efficient-lightweight YOLO: improving small object detection in YOLO for aerial images. Sens.23, 6423. 10.3390/s23146423

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 12

  KhanamR.HussainM. (2024). YOLOv11: An Overview of the Key Architectural Enhancements. 10.48550/ARXIV.2410.17725

  • CrossRef
  • Google Scholar
  • View reference in article

• 13

  Martín-RodríguezF.Álvarez-SabucedoL. M.Santos-GagoJ. M.Fernández-BarcielaM. (2024). Enhanced satellite analytics for mussel platform census using a machine-learning based approach. Electronics13, 2782. 10.3390/electronics13142782

  • CrossRef
  • Google Scholar
  • View reference in article

• 14

  MartínezG. M.CamposA. O.VilarE. G.MierR. E. L.Pérez-ArluceaM. (2011). Changes induced by mussel raft aquaculture in benthic environment of the rías baixas (galicia, Spain). J. Coast. Res. (64), 786–789. Available online at: https://www.jstor.org/stable/26482279 (Accessed August 8, 2026).

  • Google Scholar
  • View reference in article

• 15

  Mascorda CabreL.HosegoodP.AttrillM. J.BridgerD.SheehanE. V. (2021). Offshore longline mussel farms: a review of oceanographic and ecological interactions to inform future research needs, policy and management. Rev. Aquac.13, 1864–1887. 10.1111/raq.12549

  • CrossRef
  • Google Scholar
  • View reference in article

• 16

  McKindseyC. W.ArchambaultP.CallierM. D.OlivierF. (2011). Influence of suspended and off-bottom mussel culture on the sea bottom and benthic habitats: a review¹ this review is part of a virtual symposium on current topics in aquaculture of marine fish and shellfish. Can. J. Zool.89, 622–646. 10.1139/z11-037

  • CrossRef
  • Google Scholar
  • View reference in article

• 17

  McMillanC.ZhaoJ.XueB.VennellR.ZhangM. (2023). “Improving buoy detection with deep transfer learning for mussel farm automation,” in 2023 38th International Conference on Image and Vision Computing New Zealand (IVCNZ), 1–6. 10.1109/IVCNZ61134.2023.10343657

  • CrossRef
  • Google Scholar
  • View reference in article

• 18

  MoghimiA.WelzelM.CelikT.SchlurmannT. (2024). A comparative performance analysis of popular deep learning models and segment anything model (SAM) for river water segmentation in close-range remote sensing imagery. IEEE Access12, 52067–52085. 10.1109/ACCESS.2024.3385425

  • CrossRef
  • Google Scholar
  • View reference in article

• 19

  NarangG.BerardiniD.PietriniR.TassettiA. N.ManciniA.GaldelliA. (2024). “Edge-AI for buoy detection and mussel farming: a comparative study of YOLO frameworks,” in 2024 20th IEEE/ASME International Conference on Mechatronic and Embedded Systems and Applications (MESA) (Genova, Italy: IEEE), 1–8. 10.1109/MESA61532.2024.10704814

  • CrossRef
  • Google Scholar
  • View reference in article

• 20

  NASA Earth Observatory (2022). Mussel Rafts of the Ría de Arousa,” Available online at: https://visibleearth.nasa.gov/images/151746/mussel-rafts-of-the-rias-de-arousa (Accessed August 8, 2026)

  • Google Scholar
  • View reference in article

• 21

  Pérez-CamachoA.LabartaU.VinseiroV.Fernández-ReirizM. J. (2013). Mussel production management: raft culture without thinning-out. Aquaculture406–407, 172–179. 10.1016/j.aquaculture.2013.05.019

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 22

  RedmonJ.DivvalaS.GirshickR.FarhadiA. (2016). “You only look once: unified, real-time object detection,” in 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), (Las Vegas, NV, USA: IEEE), 779–788. 10.1109/CVPR.2016.91

  • CrossRef
  • Google Scholar

• 23

  RicciS. W.BohnenstiehlD. R. (2022). Monitoring visitation at North Carolina artificial reef sites using high spatiotemporal resolution PlanetScope imagery. Reg. Stud. Mar. Sci.55, 102511. 10.1016/j.rsma.2022.102511

  • CrossRef
  • Google Scholar
  • View reference in article

• 24

  RodríguezM. V. (2017). História da institucionalização dos direitos sociais: princípios e conceitos. Rev. HISTEDBR -Line17, 230. 10.20396/rho.v17i71.8649862

  • CrossRef
  • Google Scholar
  • View reference in article

• 25

  SanghaH. S.DarrM. J. (2025). Influence of model size and image augmentations on object detection in low-contrast complex background scenes. AI.6, 52. 10.3390/ai6030052

  • CrossRef
  • Google Scholar

• 26

  ShiY.YangY.Xul.SunY.YuL.DuH. (2024). “Application of YOLO11 algorithm for rotating small object detection in remote sensing images,” inproceedings of the 2024 international conference on mathematics and machine learning. (Nanjing, China: ACM), 79–84. 10.1145/3708360.3708373

  • CrossRef
  • Google Scholar

• 27

  SoliñoM.FiguerasA. (2025). The vulnerability of mussel aquaculture: understanding environmental threats and future directions. Aquaculture599, 742196. 10.1016/j.aquaculture.2025.742196

  • CrossRef
  • Google Scholar
  • View reference in article

• 28

  VargheseR.SambathM. (2024). “YOLOv8: a novel object detection algorithm with enhanced performance and robustness,” in 2024 International Conference on Advances in Data Engineering and Intelligent Computing Systems (ADICS) (Chennai, India: IEEE), 1–6. 10.1109/ADICS58448.2024.10533619

  • CrossRef
  • Google Scholar
  • View reference in article

• 29

  XuX.WangL.DengC.HeB. (2025). Pointer meter reading recognition based on YOLOv11-OBB rotated object detection. Appl. Sci.15, 7460. 10.3390/app15137460

  • CrossRef
  • Google Scholar
  • View reference in article

• 30

  YangY.YangS.ChanQ. (2025). LEAD-YOLO: a lightweight and accurate network for small object detection in autonomous driving. Sens.25, 4800. 10.3390/s25154800

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 31

  YaseenM. (2024). What Is YOLOv8: An In-Depth Exploration of the Internal Features of the Next-Generation Object Detector. 10.48550/ARXIV.2408.15857

  • CrossRef
  • Google Scholar
  • View reference in article

• 32

  YuJ.MashaijaB. A.MwankefuC. N. (2025). Binary classification of marine corrosion using deep learning: a comparative benchmark analysis across multi-source datasets. Reg. Stud. Mar. Sci.90, 104476. 10.1016/j.rsma.2025.104476

  • CrossRef
  • Google Scholar
  • View reference in article

• 33

  ZhaoF.MizunoK.TabetaS.HayamiH.FujimotoY.ShimadaT. (2024). Survey of freshwater mussels using high‐resolution acoustic imaging sonar and deep learning‐based object detection in Lake izunuma, Japan. Aquat. Conserv. Mar. Freshw. Ecosyst.34, e4040. 10.1002/aqc.4040

  • CrossRef
  • Google Scholar
  • View reference in article

• 34

  ZhouQ.DingS.FengY.QingG.HuJ. (2022). Corrosion inspection and evaluation of crane metal structure based on UAV vision. Signal Image Video Process16, 1701–1709. 10.1007/s11760-021-02126-7

  • CrossRef
  • Google Scholar
  • View reference in article

• 35

  ZhuS.AblameykoS. (2026). “AR-YOLOv11-OBB: a remote sensing oriented object detection method via adaptive rectangular convolution,” in 6th International Conference on Consumer Electronics and Computer Engineering (ICCECE). (Wuhan, China: IEEE), 81–84. 10.1109/ICCECE69169.2026.11399848

  • CrossRef
  • Google Scholar