The apple is one of the most popular fruit crops, ranking second in global fruit production. Harvesting apples remains a crucial yet demanding operation because it requires substantial labor and time (Qu et al., 2015; Jia et al., 2020). Traditional harvesting methods rely primarily on manual workforces, resulting in increased expenses, workforce shortages, and inconsistent quality and efficiency. Researchers have extensively investigated automated fruit detection and harvesting technologies that utilize machine vision and clustering-based segmentation to boost efficiency and precision (Tu et al., 2010; Jia et al., 2020).

In recent years, deep learning techniques such as YOLO, SSD, Faster R-CNN, and Mask R-CNN have been widely applied in fruit detection and recognition (Onishi et al., 2019; Biffi et al., 2020; Jia et al., 2020; Zhang et al., 2020; Xiao et al., 2023, 2023). These systems fall into two categories: single-stage models (e.g., YOLO, SSD), which directly predict object locations and classes for faster processing, and two-stage models (e.g., Faster R-CNN, Mask R-CNN), which first propose candidate regions to improve classification and bounding accuracy (Tianjing and Mhamed, 2024; Shi et al., 2025) (Likas et al., 2003; Wang et al., 2022; Mhamed et al., 2024; Tianjing and Mhamed, 2024). Recent studies have demonstrated the potential of UAV-based phenotyping and machine learning approaches for monitoring crop traits and yield in tomato and quinoa, highlighting the growing role of computer vision in precision agriculture (Johansen et al., 2019, 2020; Jiang et al., 2022a). Deep learning enhances fruit detection by extracting key colour, shape, and texture features for segmentation and recognition. However, accuracy in orchards is hindered by variable lighting, foliage cover, and clustered fruit. Moreover, reliance on large datasets, high computational demands, and long training times limits their practical use in apple harvesting (Wang et al., 2022). Moreover, they often produce only 2D bounding boxes, lacking the precise in-depth information needed for robotic harvesting. These constraints limit their suitability for real-time field deployment.

Beyond fruit detection, deep learning has advanced applications in remote sensing, radar imaging, and ecological monitoring (Guan et al., 2025). Recent studies on PolSAR ship detection (Gao et al., 2023a), scattering-aware networks, few-shot SAR classification (Gao et al., 2023b, 2024), and multi-source data fusion highlights its versatility in complex detection tasks (Shen et al., 2024; Zhang et al., 2024). These cross-domain advances reinforce the relevance of developing efficient and adaptable methods for automated fruit detection and localization.

An alternative to deep learning is clustering-based segmentation. K-Means clustering is an unsupervised learning method that groups pixels by feature similarity, enabling effective fruit segmentation under complex orchard conditions ( (Likas et al., 2003; Na et al., 2010). K-Means delivers rapid and sturdy segmentation, which stands out from other methods like Fuzzy C-Means and DBSCAN, which require more computation and struggle with noise (Song et al., 2013; Jamel and Akay, 2019; Ikotun et al., 2023). Previous studies have applied K-Means for apple recognition (Wang Dandan et al., 2015). While some researchers utilized integrated extremum methods for fruit positioning (Jia et al., 2020). Recent studies further refined segmentation with fuzzy C-means (Sarbaini et al., 2022) CNN-based semantic segmentation (Ramadhani et al., 2022; Wang et al., 2022), and monocular vision approaches (Zubair et al., 2024). However, the challenge of achieving robust performance in real orchard conditions with limited data remains (Yang et al., 2012).

This study presents an enhanced K-Means clustering segmentation algorithm combined with multi-feature fusion (colour, morphology, and texture) and stereo vision for accurate 3D localization. The approach is designed to reduce misclassification and provide depth information critical for robotic harvesting. Unlike deep learning methods, the proposed system emphasizes computational efficiency, real-time applicability, and reduced training data requirements, making it well suited to practical orchard deployment. The method is comprehensively evaluated against state-of-the-art models, including Faster R-CNN, YOLOv7, and Mask R-CNN, and demonstrates superior accuracy, reduced coordinate deviation, and stable performance across different camera angles.

The experimental setup consists of a four-arm parallel picking robot equipped with a high-precision vision system and a 3D stereo camera (1920 × 1080 pixels; Model: Hikvision MV-DL2125-04H-R) for apple detection and localization. The 3D camera was mounted at the front end of the robotic arm. Computational processing was performed on a high-performance computer running an Intel i7–12700 processor, ensuring efficient execution of clustering, segmentation, and localization tasks. Apple images were collected from a commercial orchard with diverse lighting conditions (morning, noon, evening), varying shading levels, and different apple clustering patterns to ensure a representative dataset. A dataset comprising 4,200 sample images of Aksu apples, a variety cultivated in Aksu Prefecture, Xinjiang, China, was collected. The dataset includes 2,200 images of red apples against green foliage and 2,000 images of green apples against green foliage. Each apple within the images was manually annotated using a circle-fitting method to ensure precise localization and segmentation. The dataset was split into an 8:2 ratio, with 80% used for training and 20% for testing. This choice ensured sufficient data for training while maintaining an independent set for performance evaluation. As the proposed method is based on clustering and does not require iterative hyperparameter optimization, no separate validation set was used. A similar adjustment of dataset splitting has been discussed in previous studies with small datasets (Ashtiani et al., 2021). Each image was manually annotated using LabelImg software, and apples were labelled based on their position, size, and occlusion level. To improve the model’s robustness, data augmentation was applied. Random rotation (± 15°), brightness variation (± 20%), and Gaussian noise were introduced to simulate real-world orchard variability caused by lighting changes, fruit occlusion, and viewing angle differences. This process reduced the risk of overfitting and enabled better generalization to unseen samples. Similar to findings in postharvest imaging studies (Javanmardi and Ashtiani, 2025), such augmentation strategies enhance dataset diversity and improve the reliability of classification models.

In the next section, Equations describe standard image preprocessing operations, clustering formulations, stereo vision disparity and depth estimation, and evaluation metrics are based on established methods documented in (Hartigan and Wong, 1979; Hartley and Zisserman, 2003; Gonzales and Woods, 2018). The enhanced K-means clustering and stereo vision localization method was implemented using standard Python and OpenCV libraries, with all parameters reported in this study. The dataset cannot be made publicly available due to restrictions, but a representative subset or implementation details are available from the corresponding author upon reasonable request.

The proposed clustering-based segmentation and 3D localization algorithm demonstrated consistent superiority in detection precision and spatial localization under diverse orchard conditions. Figure 5 illustrates the variation in RA and MCD under different lighting and occlusion levels. The proposed method maintained an average accuracy above 91%, while Faster R-CNN exhibited a pronounced decline when fruit overlaps exceeded 40%. In contrast, our algorithm achieved lower MCD values (≤ 0.3%), indicating more stable spatial localization across both daytime and nighttime datasets. (Figure 5). Moreover, the consistently reduced MCD values throughout all collecting distances indicate better localization accuracy of the proposed algorithm (Figures 6A, B). Figures 6C and D demonstrate that the proposed method consistently maintains a CRR above 90%, outperforming DBSCAN across varying overlap rates. The depth estimation accuracy of the stereo vision system was evaluated by comparing it with YOLOv7 and SSD across four different scenarios: complex lighting conditions, fruit occlusion, dynamic oscillation conditions, and dense target distributions. Across all four tested scenarios, the suggested model showed better recall and precision than YOLOv7 and SSD (Figure 7).

Depth estimation accuracy was further validated, achieving a maximum localization error of 0.97% across 800–1100 mm collection distances (Figure 8). Errors ranged from 0.4–0.65% at 800 mm and 0.4–0.5% at 1000 mm, with only slight increases to 0.73–0.79% at 1100 mm. All deviations remained below 1%, confirming high-precision depth estimation suitable for robotic harvesting applications. As shown in Figures 9A, B, the proposed algorithm outperformed MISA in detecting apple orientations on four trees at 0°, 45°, 90°, and 180°. It achieved the highest detection rate (> 40%) at 45°, while no apples were detected at 180°, where MISA showed greater variation and overlap, indicating reduced stability. Results for multiple algorithms at the 45° orientation are summarized in Table 1. The proposed method achieved the highest recognition accuracy (93%), correctly identifying 39 apples, followed by the CNN model (88%). The template-matching (TM) approach had the lowest accuracy (70%, 28 apples correctly identified).

In four real-world orchard scenarios, the proposed model was compared with FCN and Mask R-CNN (Figure 10). It consistently outperformed both, achieving an F1-score of 92% under varied illumination (Figure 10A) and an mAP of 91% for densely clustered fruits (Figure 10B). Under wind disturbance (Figure 10C), it maintained the highest frame rate per second (FPS), demonstrating strong real-time efficiency. Across multi-fruit orchard conditions (Figure 10D), the model again achieved the highest mAP, confirming its robustness and adaptability. Figure 11 shows that the proposed model maintained the lowest standard deviation across all camera angles (0°–45°), stabilizing after about 25 images. Even at 45°, where deviation slightly increased for all models, it remained the most stable, confirming reliable performance under varying camera orientations.

The proposed clustering-based stereo-vision approach achieved > 91% detection accuracy, < 1% localization error, and stable performance under varying lighting and camera angles, all with a modest dataset. These results demonstrate its suitability for real-time, low-cost robotic harvesting, offering reliable detection and positioning without extensive training or high computational demand—an effective solution for autonomous orchard operations in precision agriculture.

Accurate segmentation is crucial for precise apple detection in challenging orchard environments (Kang and Chen, 2020). The improved MCD and RA values indicate that multi-feature fusion with adaptive K-means clustering increases robustness to lighting changes and occlusion. Deep-learning models such as Faster R-CNN often lose accuracy under these conditions (Bargoti and Underwood, 2017; Fu et al., 2020). In contrast, the proposed unsupervised approach remains stable with fewer samples. Compared with DBSCAN, it achieved higher stability and accuracy across distances and image counts (Limwattanapibool and Arch‐int, 2017; Hartigan and Wong, 1979). These results confirm strong generalization and real-time potential for orchard use.

The success of robotic apple picking depends heavily on precise 3D localization. Our results are consistent with earlier research, where YOLO-based algorithms struggle to make real-time changes in challenging agricultural settings (Jiang et al., 2022b). This is consistent with other studies where YOLO-based models struggle in complex environments (Bresilla et al., 2019; Parvathi and Selvi, 2021). Consistent with previous studies, YOLOv7 demonstrated better accuracy and recognition speed than SSD (Wang and Chen, 2024). In contrast, a previous study showed that YOLOv7 achieved exceptional detection rates of Camellia oleifera fruit in orchards with 95.74% mAP, 93.67% F1 score, 94.21% precision, 93.13% recall and a detection time of 0.025 seconds (Wu et al., 2022). Recent research on brinjal detection using deep learning models has demonstrated the effectiveness of a lightweight YOLO architecture and edge-based computing frameworks for real-time harvesting applications (Nahiduzzaman et al., 2025; Tamilarasi et al., 2025). These approaches, while achieving high precision and recall, still depend on large, annotated datasets and relatively intensive computational resources. In contrast, our clustering-based multi-feature method achieves stable performance with fewer training samples and reduced hardware requirements, underscoring its suitability for orchard conditions. Our results are consistent with previous studies, indicating that while SSD performs well in controlled environments, it may struggle in more complex scenarios than YOLOv7. For example, Xu et al. reported lower SSD performance in typical agricultural environments where occlusions and cluttered backgrounds are very common (Xu et al., 2024). In contrast, Deng et al (Deng et al., 2024). found that YOLOv7, when enhanced with attention mechanisms, consistently outperformed SSD in citrus detection under different orchard conditions. Apple posture detection is critical in establishing the best picking strategies (Liu et al., 2024). The observed stable detection suggests that our method effectively addresses occlusion and angle-related distortions, a common challenge in fruit detection (Safari et al., 2024).

The proposed method showed stable performance relative to MISA and achieved higher accuracy than CNN, TM, and other traditional classifiers, reflecting improved feature extraction and classification capability. Similar challenges in illumination and feature consistency were also noted by (Sun et al., 2021). Consistent results under varying field conditions confirm that the model can maintain real-time reliability in orchard operations. Previous studies using FCN reported fruit-counting accuracies of 0.91–0.95 and yield accuracies up to 0.98 (Häni et al., 2020), while Faster R-CNN achieved an F1-score of 0.89 and 91% mAP. In contrast, our model achieved higher mAP, F1-score, and frame rate, demonstrating superior detection in dense, multi-fruit environments. Real-world comparison with FCN and Mask R-CNN confirmed the proposed model’s superior accuracy and processing efficiency for dense, multi-fruit environments (Wan and Goudos, 2020; He et al., 2017). Compared to previous studies, Mask R-CNN performed poorly in our study, where the precision rate reached 97.31% and the recall rate reached 95.70% (Jia et al., 2020). These outcomes highlight its stability and real-time applicability under orchard conditions. Unlike deep-learning models that rely on large annotated datasets, the algorithm maintained strong performance with limited training images, reflecting better adaptability and lower data dependence (Koirala et al., 2019). Bargoti and Underwood found that 729 training images were necessary to stabilize AP for apple detection, but almond and mango models needed more data (Bargoti and Underwood, 2017). This study also demonstrated that data augmentation enabled better apple detection using only 100 images compared to 300 images without augmentation. Similarly, 93% of apples were accurately detected in 50 images despite uneven lighting conditions in a previous study (Xu and Lv, 2018). Compared to deep learning models like Faster R-CNN and YOLOv7, the proposed method requires less computational power and no extensive training, making it suitable for real-time applications on standard hardware. While sequential processing may limit scalability in large-scale deployments, this can be optimized with parallel computing. The pipeline’s reliance on generalizable features such as colour, texture, and morphology also makes it adaptable to other fruits or crops with minor adjustments. However, large-scale field validation and integration with robotic harvesting systems are still required to confirm performance under real operating conditions, which will be addressed in future development.

In conclusion, this study presents a clustering-based stereo vision algorithm that combines K-means segmentation and multi-feature fusion for accurate apple detection and 3D localization in orchard environments. The method offers high accuracy, strong generalization, and real-time feasibility with minimal training data and computational demand—key advantages over deep-learning approaches. While sequential processing and limited field scale remain constraints, these can be addressed through parallel computing and large-scale robotic trials. Future work should focus on optimizing real-time performance and extending the framework to other fruit crops and intelligent harvesting systems.