Accurate detection of Jinxiu Malus fruits in unstructured orchard environments is hampered by frequent overlap, occlusion, and variable illumination. To address these challenges, we propose YOLOv8-MSP-PD (YOLOv8 with Multi-Scale Pyramid Fusion and Proportional Distance IoU), a lightweight model built on an enhanced YOLOv8 architecture. We replace the backbone with MobileNetV4, incorporating unified inverted bottleneck (UIB) modules and depth-wise separable convolutions for efficient feature extraction. We introduce a spatial pyramid pooling fast cross-stage partial connections (SPPFCSPC) module for multi-scale feature fusion and a modified proportional distance IoU (MPD-IoU) loss to optimize bounding-box regression. Finally, layer-adaptive magnitude pruning (LAMP) combined with knowledge distillation compresses the model while retaining performance. On our custom Jinxiu Malus dataset, YOLOv8-MSP-PD achieves a mean average precision (mAP) of 92.2% (1.6% gain over baseline), reduces floating-point operations (FLOPs) by 59.9%, and shrinks to 2.2 MB. Five-fold cross-validation confirms stability, and comparisons with Faster R-CNN and SSD demonstrate superior accuracy and efficiency. This work offers a practical vision solution for agricultural robots and guidance for lightweight detection in precision agriculture.

Non-destructive assessment of kiwifruit quality is critical for postharvest preservation and grading. This paper proposes a novel quantitative evaluation method for the kiwifruit comprehensive quality index (KCQI) during shelf life, based on hyperspectral imaging (HSI) combined with a one-dimensional convolutional neural network (1D-CNN). Hyperspectral images of two kiwifruit cultivars were acquired at four shelf-life stages using an HSI system, and six quality parameters were measured as reference standards. Based on correlation and factor analyses, five key parameters—soluble solids content, firmness, L*, b*, and chroma—were selected to construct the KCQI. Three spectral band selection methods and three modeling algorithms were compared, with the competitive adaptive reweighted sampling (CARS)–1D-CNN model yielding the highest prediction accuracy ($R_P^2$ = 0.82, RMSEP = 0.26, RPD_P = 2.39). Subsequently, a spatial distribution map was generated to visualize the KCQI. These results demonstrate the potential of the HSI–1D-CNN approach for accurate postharvest quality monitoring and intelligent grading.

To address the challenge of apple picking-point localization under severe occlusion and complex backgrounds in dwarf rootstock high-density orchards, this paper proposes a target-region-segmentation-based 3D localization method in which the fruit sphere center is defined as the picking point. A lightweight occlusion-aware semantic segmentation network, OA-LiteSegNet, is developed to generate accurate foreground masks for partially visible apples. Guided by the segmentation result, a local target-region point cloud is reconstructed from the depth map and further processed by robust sphere fitting to estimate the apple sphere center and fruit diameter. Experimental results show that OA-LiteSegNet achieves 90.35% mIoU and 76.27% BF1 with only 5.92 M parameters and 11.30 GFLOPs, while reaching an inference speed of 128.40 FPS, achieving a favorable accuracy–efficiency trade-off among the compared mainstream models. The proposed 3D localization method achieves an average sphere-center error of 2.88 mm, which is 68.4% lower than that of the conventional bounding-box-center back-projection method (9.11 mm), and maintains stable performance across different occlusion types. These results demonstrate the potential of the proposed method for real-time and accurate apple localization under controlled laboratory occlusion scenarios, while further field validation in real orchards remains necessary.

Apple-harvesting perception must remain usable when depth measurements are unavailable. This study evaluates a single four-channel YOLOv8s-OBB detector that accepts RGB-D input when depth is available and RGB-Z input, with a zero-filled fourth channel, when depth is absent. Training-only input-level modality dropout exposes the detector to both sensing states without changing the inference graph. Apples are assigned to three operational visual categories—ripe_pickable, ripe_unpickable, and unripe—and oriented bounding boxes provide image-plane geometric cues. On an image-level split of 542 orchard images, p_drop = 0.5 gave the highest observed full-test mAP₅₀ and mAP_50-95 among the tested settings, at 0.7199 and 0.5380. Under full-test forced RGB-Z input, mAP₅₀ was 0.7064, compared with 0.5906 without modality dropout and 0.6952 for the RGB-only baseline. On the paired night-time subset, mAP₅₀ curve AUCs under three graded missing-depth masks were 0.7225–0.7355, compared with 0.6088–0.6444 without dropout. These results demonstrate empirical missing-depth tolerance within a single detector under the evaluated conditions. The dropout setting is empirical, and the results are single-run point estimates. Daytime images lack paired depth, so day/night comparisons do not isolate illumination robustness; the image-level split also does not establish scene-independent generalization. OBB outputs are image-plane cues, and the exploratory geometry-only simulation does not validate robotic grasping or field harvesting success.