Biased calibration can often be found in the model parameter calibration using the traditional Gaussian likelihood function (GLF). Particularly, there are measurement errors in the observation data and complex algorithmic structures in the model. The GLF is also limited to capturing the heteroscedastic characteristics of model residuals, leading to the inaccurate calibration in Markov chain Monte Carlo (MCMC). In this study, two modified likelihood functions were introduced: the Gaussian likelihood function with coefficient of variation transformation (GLF-CV), and the Gaussian likelihood function with Box-Cox transformation (GLF-BC). These modified likelihood functions were then incorporated the heteroscedasticity induced by both the observation data and the model structure. A more accurate representation was obtained for the complex characteristics of the model residuals. The performance of the modified likelihood functions was also evaluated to conduct the parameter calibration and uncertainty analysis using the uncertainty ratio (UR). The field experimental data was collected from three ecological sites: Gaoyao Xuehuanian (early-maturing) from 2004 to 2009, Xinghua Wuyujing3 (mid-maturing) from 2001 to 2004, and Liu'an Shanyou63 (late-maturing) from 1991 to 2004. The RiceGrow and Oryza2000 phenology models were selected to compare the effects of GLF-CV, GLF-BC, and GLF on the calibration. Model parameter calibration was also conducted using the ensemble sampling for affine-invariant MCMC (EMCEE). The research findings were summarized as follows: 1) The UR ranges were 2.66-4.54, and 2.30-4.41 days, respectively, for the RiceGrow and Oryza2000 phenology models under all three likelihood functions. There were some effects of all three likelihood functions on the parameter calibration, leading to reasonable UR values within the specified ranges. 2) The GLF-BC of the RiceGrow model was achieved in the smallest UR values for the Gaoyao Xuehuanian, Xinghua Wuyujing3, and Liu'an Shanyou63 varieties. Specifically, the predicted UR values with the GLF-BC were 0.09, 0.07, and 0.80 days smaller than those with the GLF, while 1.21, 0.20, and 0.07 days smaller than those with the GLF-CV. The superior adaptability of GLF-BC was then achieved in the RiceGrow phenology model, indicating the better-improved calibration. 3) The likelihood function with the smallest UR varied greatly among the different rice varieties in the Oryza2000 phenology model. The GLF obtained the smallest UR for the Gaoyao Xuehuanian variety, while the GLF-BC and GLF-CV were the smallest UR for the Xinghua Wuyujing3 and Liu'an Shanyou 63 rice varieties, with values of 2.30, 4.17, and 3.50 days, respectively. Consequently, the likelihood function depended mainly on the primary source of heteroscedasticity in the model residuals. The optimal model was achieved in the GLF-CV, when the main source was the observation data, whereas, the GLF-BC was preferred, when the main source was the model structure. The GLF was selected as the suitable likelihood function with the minimal heteroscedasticity of model residuals. In conclusion, the heteroscedastic characteristics of model residuals were captured from the measurement errors and complex algorithmic structures. Two modified likelihood functions (GLF-CV and GLF-BC) can also provide more accurate descriptions of heteroscedasticity induced by both observation data and model structure. The comparative analysis of parameter calibration and model uncertainty with the UR demonstrated that the modified likelihood functions can be expected to effectively improve the calibration. The likelihood function also depended on the main source of heteroscedasticity, with the preferable GLF-CV under the primary source of observation data, and the preferred GLF-BC under the primary source of the model structure. The GLF was suitable for the minimal heteroscedasticity of model residuals.

A phenology prediction was proposed to improve the accuracy and generalization in soybean breeding populations in this paper. The feature disentanglement was integrated with the principle of "causal invariance". Causality-based weighted auto-encoder network (CWAE) was combined with a phenology simulation to correct prediction errors in the DSSAT-CROPGRO model. The causal weights were then calculated using the Markov boundary and partial correlation coefficient, according to the simulated features, such as the thermal and photoperiod variables, as the inputs. The low-redundancy latent features were reconstructed via a weighted auto-encoding architecture, and then progressively transferred to the subsequent phenological stages. Model validation was conducted after prediction optimization. The field trial data were collected from four critical phenological stages-flowering (R1), pod beginning (R3), seed formation(R5), and maturity initiation (R7). A diverse panel of 309 soybean genotypes was also recorded from the YangtzeHuai Soybean breeding line population (YHSBLP). Data collection occurred at the Yancheng ecological site over four phenological stages (2018–2020). The thermal and photo-period effect variables demonstrably enhanced the feature discriminability. The latent features after CWAE extraction showed significantly reduced redundancy, with a substantial reduction of 70.59%, compared with the raw inputs. The progressive feature transfer effectively improved the predictive performance over all stages, thus resulting in the reduction of the root mean square error (RMSE) by 13% to 19% for phenology predictions. Comparative analysis against the standalone CROPGRO model revealed that there was the fusion model greatly enhanced the prediction accuracy. The RMSE was reduced from a range of 4.59-5.98 days down to 3.13-4.09 days, indicating a decrease of 23.37% to 31.81%. The fusion model also demonstrated the strong generalization under environmental deployment, particularly in the 2021 year with distribution shifts. Among them, the population-level RMSE decreased by 23.53% to 71.01%, and the genotype-level average RMSE was reduced from 5.45-12.41 days to 2.80-5.00 days. Notably, the effective calibration was achieved for over 80% of the genotypes within the population under the challenging cross-environment validation. As such, the highly relevant, low-redundancy, and robust latent features were successfully extracted from the crop simulations. By enabling extraction of highly relevant, low-redundancy robust features, the CWAE framework significantly enhances phenology prediction accuracy and generalization capability in soybean breeding populations. This framework provides an effective and practical approach for analyzing phenotypic responses of soybean breeding materials to varying temperature and photoperiod conditions, based on enhanced crop growth simulations.

Grape is one of the major fruit crops in agricultural production. Manual cultivation cannot fully meet the large-scale production in recent years, due to its labor intensity, high costs, and low efficiency. Mechanical harvesting is often required for grape harvesting in the sustainable industry. Therefore, mechanization and intelligent automation can be very crucial to accelerate the cultivation. Point cloud semantic segmentation can be expected to serve as the unmanned harvesting, accurate fruit localization, and precise cutting in the context of agricultural intelligent production. However, the dense point cloud data has been widely used in the semantic segmentation in agricultural scenarios, which is difficult and costly to acquire. Sparse point cloud data can often lead to suboptimal segmentation in the practical implementation of intelligent harvesting. In this study, the point cloud semantic segmentation (called SP-Transformer) was introduced to incorporate the PointNet algorithm with a transformer multi-head self-attention mechanism. The sparse data was also selected. The SP-Transformer divided the point cloud into multi-level windows, enabling the attention mechanism to focus on the local features within each window. A multi-scale fusion strategy was employed to combine the dense and sparse features in the receptive field, and then capture the long-range contextual dependencies. Additionally, the high-level feature embedding was applied at the initial stage of the attention in order to enhance the segmentation performance of the sparse point clouds. Specifically, the three-dimensional space was first partitioned into non-overlapping cubic windows, with the points distributed across different cubic windows. Each query point was considered only the neighboring points within the same cubic window. The independent multi-head self-attention operations were performed within each window. Since the point cloud was divided into small windows, the receptive field of each query point was limited inherently. The cross-window communication was introduced to shift the window between two consecutive Transformer blocks by half the window size, thereby increasing the contextual connections for the receptive field. The input point cloud was divided into non-overlapping cubic windows using a predefined window size (WS). The data set of points (KDEN) within the same cubic window was identified in each query point (QI). Simultaneously, the input point cloud was downsampled to construct a sparse sampling window, thus dividing the downsampled space into non-overlapping cubic windows with a size of twice the WS. The set of points (KSPA) was the larger cubic window of the QI. Finally, the two sampling sets were combined to form the final sampling point set, which was used for further processing. Both local and global features were effectively captured, even in the sparse point cloud data. Experimental results on a grape dataset demonstrated the effectiveness of the SP-Transformer, with an average accuracy of 89.9%, which was 4.4 percentage points higher than PointNet++ and 1.5 percentage points higher than Point Transformer. The segmentation accuracy reached 81.1% for the grape classes, with an average intersection over union (IoU) of 82.8%. To validate the robustness of the model, the point cloud completion algorithm (PENet) was employed to densify the point cloud. The point cloud density increased from approximately 60 000 points to around 500 000 points. A comparison was performed on the dense and sparse data. The SP-Transformer was trained to evaluate the real-time performance and segmentation accuracy. The results indicate that the SP-Transformer maintained the robust segmentation performance, even for the low-density point clouds, thus outperforming traditional PointNet++ in terms of both accuracy and efficiency. The SP-Transformer represented a significant advancement in the point cloud semantic segmentation, particularly in the agricultural applications where sparse point cloud data was prevalent. The multi-head self-attention mechanism and multi-scale fusion strategy effectively solved the challenges posed by sparse data, enabling more accurate and efficient segmentation. As such, the intelligent harvesting was enhanced to accurately locate and segment the fruits, thereby reducing the labor costs for high operational efficiency. Moreover, the SP-Transformer also shared the broad implications beyond agricultural applications, particularly on the sparse point cloud data in autonomous driving, robotics, and environmental monitoring. The cross-window communication and high-level feature embedding can provide a robust framework to capture both local and global features. A versatile tool can also be used in various tasks of computer vision. In conclusion, the SP-Transformer can be expected to serve as a powerful and efficient solution for the point cloud semantic segmentation in agricultural intelligent harvesting. The sparse data can be utilized to capture the long-range contextual dependencies for more accurate and reliable intelligent harvesting. The high accuracy of the segmentation on the sparse point cloud data can greatly contribute to the advancement of computer vision in the field of precision agriculture.

Accurate and rapid estimation of spatial posture is crucial to monitoring the behavior in group-housed livestock. Among them, 3D posture estimation can offer precise spatial data in the conditions of occlusion, compared with the traditional 2D. The current techniques of 3D posture estimation are primarily applied in human and autonomous driving fields, thus depending mainly on expensive measurement equipment and datasets. However, it is still challenging for animal behavior and management at present. Therefore, it is urgently needed for the low-cost and efficient measurement of animal behavior posture. In this study, a general approach was proposed to estimate the 3D posture of livestock using binocular stereo matching. Firstly, a modified model of stereo matching was employed to obtain the depth information using deep learning. Then, a top-down 2D posture model was used to extract the target bounding boxes and then detect the key points. Finally, the locations of key points were mapped back to the image space and then fused with the stereo-matching model for the 3D posture information. Since the matching accuracy depended on the precise depth information, the main challenges in stereo matching were attributed to the thin structure and weak texture matching. ACLNet stereo matching model was constructed using attention mechanism and ConvGRU iterative refinement. The relative depth layers of image textures were encoded to restrict the attention of the model to the areas near the true disparity. The high-precision depth information was gradually recovered in a residual manner. Ablation experiments and generalization tests were carried out on the Scene Flow dataset and the Middlebury dataset, respectively, in order to validate the effectiveness of the ACLNet model. The results show that the ACLNet model was achieved in an endpoint error (EPE) of 0.45 on the Scene Flow dataset, with a reduction of 0.37 pixels, compared with the baseline model without attention and ConvGRU mechanisms. Better generalization was also performed on real-world datasets, such as Middlebury. The EPE was 0.56 on the goat depth dataset; The mean per joint position error (MPJPE) on the goat 3D posture test set reached 45.7 mm in the improved model. There was a decrease of 21.1 mm, compared with the baseline. Strong generalization and versatility were obtained to accurately estimate the livestock 3D posture without additional training. The 3D posture estimation experiments were also verified to take the goats as test subjects. Binocular images were only required to accurately obtain the 3D posture. The feasibility of high-precision 3D posture estimation on livestock was then validated using a simple binocular system. The finding can provide a viable solution to accurate 3D posture estimation using low-cost stereo cameras.