Real-time monitoring of flat peaches is often needed during different growth stages under natural orchard environments, particularly for yield estimation, precise fertilization, and mechanical harvesting. However, conventional approaches can only detect peaches at a single specific stage, showing low adaptability during different growth stages of the same fruit. Accurate identification is also limited to the severe occlusion from peach branches and foliage, with dense overlapping clusters of young fruit and significant variations in size, color, fuzziness, and texture during growth, particularly during the color change and enlargement stages. In this study, an Orchard Growth Detection Network (OGD-Net) was proposed using the lightweight YOLOv11n model. Sufficient accuracy was also obtained to identify flat peach growth stages within natural orchard environments. First, an improved adaptive orchard downsampling module (AOD-Module) was designed to integrate the context guided block (CG Block) with depthwise separable convolution to replace the original downsampling layer in the backbone network, enhancing peach feature expression with a low computational cost. Second, an orchard gated spectrum block (OGS-Block) was proposed to combine frequency-domain dynamic convolution with a gating mechanism. The original C3K2 module was replaced for dual frequency-spatial perception and precise identification at different growth stages. Third, an orchard-adaptive feature fusion pyramid (OAFF-Pyramid) was constructed in the neck network. Within OAFF-Pyramid, an orchard attentive upsampler (OAS-Sampler) was designed to combine EUCB and the Shift Channel Mix mechanism to reconstruct fruit contours occluded by branches and leaves; An OAF-Block was designed to fuse C2f with MSCB, CSP, and Shift Channel Mix to replace the original C3k2 feature extraction module. Finally, a separated and enhanced attention module (SEAM) was introduced at the detection head to compensate for feature loss caused by occlusion and model robustness under branch and leaf occlusion conditions. A flat peach dataset was constructed to collect 4 807 original images at the 143rd Regiment flat peach orchard in Shihezi, Xinjiang, China, covering three growth stages, namely the green and hard stage, the expansion and color transition stage, and the mature stage. A final dataset of 11 535 images was obtained after data augmentation was applied only to the training set. Experimental results demonstrated that: 1) The OGD-Net model achieved 6.0 G floating-point operations and a model size of 4.7 MB. On the test dataset from three different growth stages of flat peaches, the better performance was achieved, with the precision, recall, F₁ score, mAP_0.5, mAP_0.75, and mAP_0.5-0.95 of 88.17%, 84.60%, 86.31%, 93.14%, 89.12%, and 85.48%, respectively, which were improved by 0.75, 3.25, 2.04, 3.28, 5.36, and 6.38 percentage points over YOLOv11n. 2) Compared with YOLOv11n, the floating-point computation volume and model size of OGD-Net were reduced by 4.76% and 9.62%, respectively, indicating excellent lightweight suitability for mobile-terminal deployment. 3) In fruit enlargement and color transition, the detection accuracy mAP_0.5 reached 89.63%, which improved by a 4.51 percentage point over YOLOv11n. 4) Compared with classic neck network architectures, such as BiFPN, GFPN, Gold-YOLO, and CGRFPN, the OAFF-Pyramid achieved an R-value of 83.97% and mAP_0.5 of 90.72%, with a model size of only 4.5 MB, which was 13.46% smaller than the baseline. The heatmap highlights were highly concentrated on key characteristic areas of peaches, with no false detections or background false responses observed, indicating that the OAFF-Pyramid architecture exhibited superior detection and stronger generalization for growing peaches in complex orchard environments. 5) OGD-Net was deployed on an MMX5 Plus robot to validate the practicality of the model. Recognition success rates of 100%, 100%, and 93.3% were achieved under no occlusion, branch and leaf occlusion, and fruit overlap conditions, respectively, using an indoor peach orchard simulation platform. Grasping success rates of 93.3%, 86.7%, and 80% were obtained for mature peaches. The OGD-Net model achieved accurate identification of flat peaches during the entire growth period in natural orchard environments. The findings can provide the technical support to monitor precise growth and mechanical harvesting during orchard production.

Hami melon has been one of the most favorite fruits in Xinjiang of Northwest China. In this study, a portable spectral detection device was developed for the on-site non-destructive testing on the soluble solids content (SSC) in the hami melons. The dominant varieties of hami melons were selected as Jinsenianhua, Huangjinmi, and Xizhoumi25. A general prediction model was then constructed for the mixed varieties. A diffuse transmission spectroscopy was also integrated with the interchangeable light source and a spectroscopy acquisition module. A systematic analysis was implemented to clarify the spectral response of the internal quality of hami melons under different detection postures and light source powers. The spectroscopy acquisition scheme was then optimized to design the light source arrangement and spectroscopy acquisition module. The stability and accuracy of the spectral acquisition were improved for the diffuse transmission spectroscopy acquisition of hami melons. Spectra data was then collected at 90° intervals along the equatorial part of each hami melon. The averaged spectra from each position were fused as the original input data. Five spectral pre-processing approaches were adopted to improve the data quality: detrending (DT), standard normal variate (SNV), multiplicative scatter correction (MSC), Savitzky-Golay smoothing (SG), and first derivative (FD). The background noise was reduced to correct the baseline drift. The scattering artifacts were removed after modification. According to the pre-processed spectra, four algorithms of the characteristic wavelength selection—competitive adaptive reweighted sampling (CARS), successive projections algorithm (SPA), uninformative variables elimination (UVE), and synergistic interval partial least squares (SiPLS) were adopted to screen the variables of the pre-processed spectra, and then extract the key wavelengths. A general prediction model was constructed for the SSC of hami melons. Three modeling approaches were established after prediction, namely partial least squares regression (PLSR), support vector regression (SVR), and one-dimensional convolutional neural network(1D-CNN). The results showed that the FD-UVE-1D-CNN model was achieved the best prediction performance. The correlation coefficients between the correction and the prediction set (R_C and R_P) were 0.913 and 0.896, respectively, while the root mean square errors (RMSEC and RMSEP) were 0.700% and 0.670%, respectively, and the residual prediction deviation (RPD) reached 2.256. An improved one-dimensional residual convolutional neural network (1D-Res-CNN) model was performed the feature fusion on the convolutional layers. A residual module was introduced to further improve the prediction accuracy and robustness of the model. Furthermore, the performance of this model was significantly improved after FD-UVE pre-processing, compared with the traditional 1D-CNN. Its RC and RP increased to 0.947 and 0.923, respectively, while the RMSEC and RMSEP decreased to 0.550% and 0.581%, respectively. Compared with the traditional model, the RC and RP increased by 3.7% and 3.0%, respectively, while RMSEC and RMSEP decreased by 21.4% and 13.3%, respectively. The RPD also increased to 2.601, with an increase of 15.3%. The accuracy and stability of the general model and non-destructive testing device were verified to predict the SSC in the different varieties of hami melons. The general prediction model was deployed into the device control module, particularly for the FD-UVE-1D-Res-CNN mixed variety. Ten Jinsenianhua, ten Huangjinmi and ten Xizhoumi25 were selected without modelling before. The coefficient of variation and residual analysis were conducted on the SSC prediction of the three varieties. The average coefficients of the variation of Xizhoumi25, Jinsenianhua and Huangjinmi were 2.204%, 2.794%, and 2.096%, respectively, with the maximum absolute residuals of 0.60%, 0.50%, and 0.67%, respectively. The coefficient of variation of the SSC predictions were lower than 4.5% for all varieties, and the prediction residuals were mainly concentrated within ±0.7%, indicating the cross-variety generalization and robustness. The SCC non-destructive testing device for the hami melons was effectively treated with the spectral differences among varieties and environmental noise interference, fully meeting the requirements for the prediction stability and accuracy in field applications. The better portability and field application of the device were achieved, compared with the traditional experimental platforms and large-scale detection. The overall dimensions of the machine were 300 mm×240 mm×150 mm, with a weight of 3.6 kg. A compact structure was integrated with a power module to uniformly supply power for the main control unit, light source and heat dissipation. The stable operation was equipped with the "Hami melon SSC detection system" software, which was integrated data collection, analysis, prediction and data output functions. The plug-and-play was supported for only 5 s to detect a single fruit. This device was suitable for various on-site detection scenarios, such as orchards, markets and sorting centers. In conclusion, the hardware-software solution was integrated for the rapid, non-destructive SSC detection in hami melons. Advanced spectral pre-processing was combined with the feature screening, machine learning and deep learning. The modelling accuracy and stability were improved significantly after processing. The portable detection device with the universal model of mixed varieties can meet the on-site rapid detection requirements in the various varieties of hami melons. The finding can provide a strong reference and technical support to the fruit quality and non-destructive detection of the internal quality of the fruits.

Defects are required to be removed in many rows of dried Hami jujubes. In this study, an online detection and counting system was designed for dried Hami jujubes that were randomly distributed on the conveyor belt. The system was divided into four parts, including a target detector, multi-target tracker, counter, and software interface. The target detector was used to detect the defect categories of dried Hami jujube and then mark the detection box. The multi-target tracker was used to distribute the identity information of dried Hami jujube in the video. The counter was used to count the number of dried Hami jujube in each defect category. The software interface was used to control the program execution. The accuracy of multi-target tracking depended mainly on target detection. The target detector was first designed to compare several models of YOLO series, in order to better detect the defects with high counting accuracy. The better detection of defects was achieved in the YOLOv7 model. After that, the 3×3 CBS convolution layer in the YOLOv7 model was replaced with DCNv2, DWconv, PConv, GSconv, and DSconv. A comparison was made to explore the influence of the improved model on the defect detection of dried Hami jujube. It was found that the DSconv convolution of the improved YOLOv7 model shared fewer parameters, the higher detection speed, and the better accuracy, compared with the original. Four attention mechanisms were compared, including CBAM, ECA, SE, and SimAM. ELAN and SPP layers were also added to the backbone network of the YOLOv7 model. SimAM was performed the best to increase the accuracy of target detection. The ByteTrack multi-target tracker was selected to deal with the high similarity in the different kinds of dried Hami jujubes under the actual environment. The reason was that the appearance information dominated the performance of the multi-target detector for data association. A line-drawing counting was also proposed for statistical counting. The coordinates and ID information were focused at both ends of the counting line to count. Finally, the counting performance was achieved for dried Hami jujube. The software interface was designed using PyQt5 in the practical application. The model was then verified on the detection system of dried Hami jujube. The improved model was also deployed into the detection and counting system for online detection and counting of dried Hami jujube. The experimental results show that: 1) The floating-point computation of the improved YOLOv7 model was 64.6 G, which was 38.6% lower than that of the YOLOv7 model. The mAP, Recall, and F₁-score on the test set reached 98.03%, 93.43%, and 95.00%, respectively, which increased by 4.40, 6.88, and 7.00 percent points, respectively, compared with the original. 2) The counting model with ByteTrack as the target detector showed an accuracy rate of 90.12% for the multi-category counting. The finding can also provide technical support for the detection, counting, sorting, and grading of dried Hami jujube.

Each organ can effectively represent the health degree of cotton. It is also required for the physiological and phenotypic analysis of each organ, in order to monitor the growth and development status of cotton plants. However, the three-dimensional point cloud makes it difficult to segment newborn leaf organs at the top of the cotton seedling stage. In this study, a newborn leaf segmentation was proposed to combine Laplacian skeleton extraction and Quickshift++ using a distance field. The five steps mainly included: skeleton extraction, skeleton decomposition, new leaf extraction, new leaf organ segmentation, and phenotype parameter extraction. In skeleton extraction, the Laplace algorithm was used to extract the skeleton of a cotton plant. In skeleton decomposition, the plant skeleton was decomposed into stem and partial leaf subskeleton, according to the morphological characteristics of the cotton plant. The coordinate system of plant space was then established. The Z-axis of the coordinate system was overlapped with the stem axis, in order to better represent the typical morphology of cotton plants. The point cloud of the cotton plant was transformed from the original to the spatial coordinate system of the cotton plant. Subsequently, the skeleton decomposition was carried out in the spatial coordinate system of the cotton plant. A distance field of organ segmentation was also proposed to accurately segment newborn leaf organs using Quickshift++. The reason was that the Laplacian algorithm was limited to extracting new leaf skeletons, easily leading to over-segmenting newborn leaves, especially for the cotton plants with newborn leaves and the very small distance among them. As such, the whole organ segmentation of the cotton plant was realized at the seedling stage. The global spatial features of the newborn leaf were also encoded to rapidly locate the precise segmentation of the organ. In the segmentation of newborn leaves, the Quickshift++ algorithm with distance field as the input was first used to extract the tip point clouds of different organs of newborn leaves. Secondly, the neophylla apex cloud was identified using the local geometric features. Then the stem point cloud of newborn leaves was segmented to combine with the median normalized vector growth segmentation. Finally, the Quickshift++ algorithm was used to segment the point cloud of newborn leaf organs. The results showed that the average accuracy of organ segmentation was 0.977, the average recall rate was 0.978, the average micro F1 score was 0.976, and the average overall accuracy was 0.983. The determination coefficients of stem height, leaf width, leaf length, and leaf area were 0.982, 0.962, 0.971, and 0.949, respectively. The root mean square errors were 0.119 cm, 0.182 cm, 0.163 cm, and 1.458 cm², respectively. The improved model can accurately segment the cotton plant organs at the seedling stage. The findings can also provide an effective technical means for the high-throughput phenotypic detection of cotton.