Search

• Early Candidate Elimination Module reduces computational load. • The Adaptive Appearance Update enhances the reliability of model updates. • The model exhibits good generalization capabilities. • This study provides technical support for motion analysis of dairy goat. Intelligent tracking and monitoring are fundamental to achieving precise posture recognition, health assessment, and accurate breeding in dairy goats. In view of the motion characteristics of dairy goats, which often exhibit significant posture changes due to their lively and active movements, we present an object tracking method for a single dairy goat based on MixFormer optimized through adaptive token removal and efficient appearance update. Firstly, we collected motion data from two different breeding farms and trained the model solely on one farm and tested it in the other to verify its generalization performance. Then, through an early Candidate Elimination (CE) module, we pre-filter redundant background and similar object information in dairy goats, thereby reducing computational load and enhancing tracking speed. Finally, we introduce the adaptive appearance update strategy named Average Peak-to-Correlation Energy (APCE) to maximize the utilization of target-related historical information, thus improving target localization accuracy. The experimental results indicate that the improved MixFormer achieves 75.79% and 74.28% in terms of AUC (Area Under Curve) and Precision, respectively. These results represent a 2.75% and 7.38% enhancement over the MixFormer, outperforming KYS, KeepTrack, STARK and TOMP. The improved MixFormer runs at 30.5 fps, meeting real-time tracking requirements. This study demonstrates that the proposed approach offers an effective solution for automatic tracking and monitoring of dairy goats in complex breeding conditions. [ABSTRACT FROM AUTHOR]

• HorBlock-decoupled head more accurately locates and completely model missing stolon. • Stem Block module increases feature expression ability and unifies feature channels. • Stolon-YOLO among all methods achieves the best performance. • The study provides a technical reference for related stolon detection research. Stolons are an essential nutritional organ of strawberry seedlings, whose quantity and vigorous growth directly impact the quality of the seedling cultivation. Accurately detecting stolon is essential for increasing productivity in seedling factories. To alleviate these issues of low recognition accuracy of stolon caused by complex backgrounds, diverse growth types and low proportion of effective information in the bounding box, we propose a Stolon-YOLO for visual recognition of stolon for the first time. In our method, improvements are focused on two aspects: HorBlock-decoupled head and Stem Block feature enhancement module. Firstly, a new decoupled head is designed by introducing the HorBlock, which enables the interaction of high-order spatial information and enhances the precision for locating stolons. Sequential gnConv of the HorBlock realizes better global modeling for incomplete and curved stolon to improve detection efficiency. Then, the Stem Block is used to reinforce feature expression and reduce dimensionality. We fuse the processing results of the Stem Block with the output of HorBlock in a residual form. Experimental results demonstrate that the Stolon-YOLO achieves 92.5% in precision, 89.7% in recall rate, 91.1% in F1 score and 88.5% in average precision for stolon detection, which outperforms the standard YOLOv7 by 3% in recall rate and 3.4% in average precision. Compared with YOLOX, SSD, CenterNet and Faster R-CNN, the Stolon-YOLO increases recall rate by 8.3%, 27.4%, 17%, and 3.5%, and increases AP by 12.2%, 28.9%, 2.4%, and 7.5%, respectively. Meanwhile, the Stolon-YOLO achieves a frame rate of 50.25 FPS, meeting real-time detection demands. The above results highlight that this study provides a fast and effective method for identifying the stolon of strawberry seedlings in glass greenhouses. [ABSTRACT FROM AUTHOR]

• The fusion module reuse Transformer features to save computing resources. • Asymmetric fusion strategy improves the accuracy of recognition model. • Utilizing Kullback Leibler divergence to generate a low-noise goat template. • This work provides a high-precision method for accurately tracking dairy goat. As an important task in precision breeding, accurately tracking dairy goats is essential for their disease prediction and abnormal behavior recognition. For efficiently capturing local features in Transformer and reducing the noise in the template of dairy goat, we develop an innovative dairy goat tracking algorithm that introduces an asymmetric fusion architecture and Kullback Leibler loss to improve accuracy and generate a low-noise template. Based on TrDiMP, we construct a fusion mechanism that embeds high-dimensional timing information into its evaluation branch to enhance discrimination ability. Then, we exploit an asymmetric attention mechanism to incorporate convolutional features of different sizes and raise the percentage of local linear information in smaller features. Finally, we incorporate Kullback Leibler divergence to train the loss between current frames and the goat template, guiding the template to conform to the probability density distribution of dairy goats and reducing noise. The AUC (Area Under Curve) and Precision of the improved TrDiMP reach 77.56% and 66.10%, respectively, which are 2.91% and 4.55% higher than that of TrDiMP. The average tracking speed of the improved TrDiMP is 14.90 FPS (frames per second), 0.36 FPS faster than that of TrDiMP. The overall results demonstrate that our model can track dairy goats and provide an effective approach to capture their dynamic movement in a real pasture. [ABSTRACT FROM AUTHOR]

• The integral constraint between fruit and stem eliminates the interference of branch. • Multi-classification prediction subnetwork with the balanced loss accurately detects fruit and stem. • Adaptive feature pooling fuses the bottom features and enhances the segmentation accuracy of stem. • This work provides a path for accurately locating cherry tomatoes in plant factory. Accurate recognition of cherry tomatoes is a key issue for the automatic picking system in plant factories, which helps to improve picking efficiency and reduce production costs. By using the depth information and considering the prior adjacent constraint between the fruit and the stem of cherry tomatoes, this paper proposes an improved Mask R-CNN for visual recognition of cherry tomatoes. Firstly, the input layer of the network is modified to achieve dual-mode data fusion of RGB and depth images. Secondly, by constructing the corresponding region generation network to indicate the integral constraint between the fruit and the stem, false recognition of branches is reduced. Thirdly, a multi-class prediction subnetwork is used to decouple the pixel-level category predictions of fruit and stem. Meanwhile, multi-task loss balance and adaptive feature pooling are adopted to overcome the limitation caused by the size difference between fruit and stem. The experimental results show that the improved Mask R-CNN achieved an accuracy of 93.76% for fruit recognition, which is 11.53% and 15.5% higher than that of the standard Mask R-CNN and YOLACT, and it achieves an accuracy of 89.34% for stem recognition, which is 13.91% and 19.7% higher than that of the standard Mask R-CNN and YOLACT, respectively. Besides, the recall rate of the proposed method for stem recognition is 94.47%, which is 11.53% and 8.3% higher than that of YOLACT and Mask R-CNN, respectively. In addition, the proposed method takes only 0.04 s to process a single image, providing an efficient approach for automatically picking cherry tomatoes in plant factories. [ABSTRACT FROM AUTHOR]

[Display omitted] • Different window sizes will affect the monitoring accuracy of soil moisture; • It is feasible to select the optimal window size with the local variogram; • The window size of 13 * 13 is the optimal window size applicable in the study area. Soil moisture content monitoring with UAV remote sensing always involves the selection of an appropriate window size for spectral information extraction, but research on the effect of window size on the accuracy of soil moisture content monitoring models and the selection of the optimal window size has been rarely reported. To solve these problems, an experiment was conducted on three typical bare plots in Shahaoqu Experimental Station at Hetao Irrigation District, Inner Mongolia, China. First, remote sensing images were obtained from the three bare plots from April 15 through 17, 2019, with a six-rotor UAV equipped with a six-channel multispectral camera. Synchronously, the moisture content at 0–10 cm of the surface soil was measured using the drying method. Then, the spectral information was extracted through windows of 16 different sizes (ranging from 1 * 1 to 31 * 31). Followed was the construction of thirty spectral indices using the ratio and normalized ratio methods, and the processing of the constructed indices using principal component analysis. The principal components accounting for 95% of the cumulative contribution rate were selected as the input variables for the construction of the monitoring models based on BP neural network. Finally, the model accuracy was tested using ANOVA, and the local variogram of the spectrum was used to explore the optimal window size selection. The results demonstrated: (1) There are differences in the spectral information extracted from different sizes of windows, which affects the accuracy of soil moisture monitoring model; (2) The spatial autocorrelation threshold of the plots at the local variogram was 13 * 13, resembling the window size with the highest accuracy, so it is feasible to select the optimal window size with the local variogram; (3) As the window size of spectral information increased, R2 first increased and then decreased, reaching the maximum value of 0.261 at the size of 13 * 13, and RMSE first decreased and then increased, reaching the minimum value of 0.017 when at the size of 7 * 7. These results can provide some reference for window size selection in spectral information extraction to monitor soil moisture content. [ABSTRACT FROM AUTHOR]

• Fractional-order derivative algorithm was used to process soil VIS-NIR spectroscopy. • The correlation coefficient with most soil salt ions were best at 0.80–0.90 order. • Three-dimensional indices were used to magnify the soil salinity information. • Iteratively retaining informative variables algorithm had the best effect. • Mg2+ showed the highest estimation accuracy, with an RPD and RPIQ up to 3.81 and 3.71 respectively. This study aimed to improve the potential of visible-near infrared (VIS-NIR) spectroscopy in predicting the contents of salt and major soluble ions in the topsoil at Hetao Irrigation District in Inner Mongolia, China. A total of 120 soil samples (sampling depth is 0–20 cm) were collected from the field. Fractional-order derivatives (FODs) (with intervals of 0.05 and range of 0–2) were used for soil spectral pretreatment. Coefficient of determination (R2) was used to determine the optimal FOD spectrums of the soil salt and major soluble ions. Extreme learning machine (ELM) models were calibrated with such spectral parameters as band reflectance, newly constructed two-band indices, and three-band indices from the optimal FOD spectrum, and the models were subsequently applied to estimate the contents of soil salt and soluble ions. Different algorithms including Monte Carlo uninformative variable elimination (MCUVE), iteratively retaining informative variables (IRIV), and bootstrapping soft shrinkage (BOSS) were applied to variable selection. The results indicated that the correlation between the contents of soil salt and soluble ions and FOD spectra increased as the derivative order increased. The optimal FOD spectrum of soil salt content (SSC), K+, Na+ and SO 4 2− were the 0.9-order derivative spectrum while the optimal FOD spectrum of Ca2+, Mg2+, CO 3 2–, HCO 3 – and Cl− were 1.65, 0.85, 1.90, 0.80 and 1.55 order derivative spectrum, respectively. The newly proposed three-band index (TBI4) showed a better correlation with ions content (e.g., TBI4 of Mg2+ had the best effect, and its R2 was 0.92). In most cases, the IRIV-ELM models outperformed MCUVE-ELM and BOSS-ELM models in estimating the contents of soil salt and soluble ions. Difference existed among the optimal prediction accuracy of the contents of soil salt and different ions: the prediction accuracy of SSC, Ca2+, Na+, Mg2+, Cl− and SO 4 2− was very high while that of CO 3 2− was high and K+ and HCO 3 2− was relatively lower. However, these models have reached and even exceeded the quantitative prediction level. Therefore, we concluded that the contents of soil salt and soluble ions could be estimated using FOD algorithm and the optimal combination of spectral indices. [ABSTRACT FROM AUTHOR]

• High-resolution fused multispectral image conduce to recognize lodging sunflower. • Deep learning outperforms the classical SVM in lodging recognition accuracy. • Improved SegNet achieves the best performance in accuracy and generalization. • It is useful for the sunflower lodging hazard assessment and harvesting survey. Sunflower lodging is a common agricultural disorder taking place in the middle and late sunflower growth periods. This disorder reduces the sunflower seed yield, damages the seed quality, and hence usually causes great losses in both crop quantity and quality. Sunflower lodging is mainly caused by extreme and destructive weather events, which have been recently occurring more frequently. This is why it is highly crucial to develop methods for fast and accurate identification of sunflower lodging. In this work, an efficient method for sunflower lodging identification is proposed based on image fusion and deep semantic segmentation of remote sensing images obtained from an unmanned aerial vehicle (UAV). First, the resolution of low-resolution multispectral images was enhanced through matching their features with those of high-resolution visible-range images. Then, for effective lodging assessment, high-quality multispectral images with rich spectral information and high spatial resolution were obtained through fusing the visible-range images and the enhanced multispectral ones. Subsequently, in order to refine the identification outcomes, a variant of the segmentation network (SegNet) deep architecture was developed for semantic segmentation. This variant has skip connections, separable convolution, and a conditional random field. Experimental evaluation shows that the fusion-based approaches clearly outperform the no-fusion ones in terms of the lodging identification accuracy for all compared architectures including support vector machine (SVM), fully convolutional network (FCN), SegNet, and the proposed SegNet variant. Meanwhile, the deep semantic segmentation methods consistently outperform the classical SVM one with hand-crafted features. As well, the improved SegNet method outperformed all of the compared methods and achieved the best accuracies of 84.4% and 89.8% without and with image fusion, respectively, on one test. The corresponding accuracies on another test set were 76.6% and 83.3%, respectively. Moreover, the proposed method can also identify the sunflower lodging and non-lodging patterns and separate them from the background. These capabilities are highly beneficial for lodging hazard assessment and sunflower harvest survey. Overall, the proposed method effectively exploited UAV remote sensing image data with fusion and deep semantic segmentation modules in order to provide a useful reference for sunflower lodging assessment and mapping. [ABSTRACT FROM AUTHOR]