Computer vision provides a real-time, non-destructive, and indirect way of horticultural crop yield estimation. Deep learning helps improve horticultural crop yield estimation accuracy. However, the accuracy of current estimation models based on RGB (red, green, blue) images does not meet the standard of a soft sensor. Through enriching more data and improving the RGB estimation model structure of convolutional neural networks (CNNs), this paper increased the coefficient of determination (R²) by 0.0284 and decreased the normalized root mean squared error (NRMSE) by 0.0575. After introducing a novel loss function mean squared percentage error (MSPE) that emphasizes the mean absolute percentage error (MAPE), the MAPE decreased by 7.58%. This paper develops a lettuce fresh weight estimation method through the multi-modal fusion of RGB and depth (RGB-D) images. With the multimodal fusion based on calibrated RGB and depth images, R² increased by 0.0221, NRMSE decreased by 0.0427, and MAPE decreased by 3.99%. With the novel loss function, MAPE further decreased by 1.27%. A MAPE of 8.47% helps to develop a soft sensor for lettuce fresh weight estimation.

A Chinese solar greenhouse (CSG) is an agricultural facility type with Chinese characteristics. It can effectively utilize solar energy during low-temperature seasons in alpine regions. The low construction and operation costs make it a main facility for agricultural production in the northern regions of China. It plays an extremely important role in “Chinese vegetable basket projects”. Energy is one of the important issues faced by CSGs. The better climate resources in the northern regions of China make it possible to apply solar energy as a green and sustainable energy source in the production of CSGs. Faced with the increasingly serious environmental problems of the new century, the Chinese government has made a decision and put it into practice to improve the rate and efficiency of solar energy utilization in agricultural facilities. In this paper, we summarize the research on the application of photovoltaic power generation and solar thermal technology in CSGs. The application of these advanced solar technologies has made great progress. With the further improvement of economic benefits and the establishment of relevant support policies and incentive mechanisms, the combination of CSG and solar energy technology will have certain application prospects and satisfy China’s requirements for long-term sustainable development.

Due to the heavy computation load of closed-loop simulations, optimal control of greenhouse climate is usually simulated in an open-loop form to produce control strategies and profit indicators. Open-loop simulations assume the model, measurements, and predictions to be perfect, resulting in too-idealistic indicators. The method of two-time-scale decomposition reduces the computation load, thus facilitating the online implementation of optimal control algorithms. However, the computation time of nonlinear dynamic programming is seldom considered in closed-loop simulations. This paper develops a two-time-scale decomposed closed-loop optimal control algorithm that involves the computation time. The obtained simulation results are closer to reality since it considers the time delay in the implementation. With this algorithm, optimal control of Venlo greenhouse lettuce cultivation is investigated in Lhasa. Results show that compared with open-loop simulations, the corrections in yield and profit indicators can be up to 2.38 kg m⁻² and 11.01 CNY m⁻², respectively, through closed-loop simulations without considering the computation time. When involving the time delay caused by the computation time, further corrections in yield and profit indicators can be up to 0.1 kg m⁻² and 0.87 CNY m⁻², respectively. These conservative indicators help investors make wiser decisions before cultivation. Moreover, control inputs and greenhouse climate states are within their bounds most of the time during closed-loop simulations. This verifies that the developed algorithm can be implemented in real time.

To address the low estimation accuracy of deep learning-based crop yield image recognition methods under untrained shooting distances, this study proposes a shooting distance adaptive crop yield estimation method by fusing RGB and depth image information through multi-modal data fusion. Taking strawberry fruit fresh weight as an example, RGB and depth image data of 348 strawberries were collected at nine heights ranging from 70 to 115 cm. First, based on RGB images and shooting height information, a single-modal crop yield estimation model was developed by training a convolutional neural network (CNN) after cropping strawberry fruit images using the relative area conversion method. Second, the height information was expanded into a data matrix matching the RGB image dimensions, and multi-modal fusion models were investigated through input-layer and output-layer fusion strategies. Finally, two additional approaches were explored: direct fusion of RGB and depth images, and extraction of average shooting height from depth images for estimation. The models were tested at two untrained heights (80 cm and 100 cm). Results showed that when using only RGB images and height information, the relative area conversion method achieved the highest accuracy, with R² values of 0.9212 and 0.9304, normalized root mean square error (NRMSE) of 0.0866 and 0.0814, and mean absolute percentage error (MAPE) of 0.0696 and 0.0660 at the two untrained heights. By further incorporating depth data, the highest accuracy was achieved through input-layer fusion of RGB images with extracted average height from depth images, improving R² to 0.9475 and 0.9384, reducing NRMSE to 0.0707 and 0.0766, and lowering MAPE to 0.0591 and 0.0610. Validation using a developed shooting distance adaptive crop yield estimation platform at two random heights yielded MAPE values of 0.0813 and 0.0593. This model enables adaptive crop yield estimation across varying shooting distances, significantly enhancing accuracy under untrained conditions.

Non-destructive estimation of leafy vegetable fresh weight is crucial for precision management in both greenhouse and open-field production. However, mutual occlusion between plants in dense canopies poses a significant challenge to image-based estimation accuracy. This study systematically investigates the potential of deep learning-based image inpainting methods to reconstruct occluded regions in RGB lettuce images, thereby improving input data quality for downstream weight estimation models. Three state-of-the-art inpainting models—Vision Transformer-based Denoising Autoencoder (ViT-DAE), Aggregated Contextual–Transformation Generative Adversarial Network (AOT-GAN), and a conditional Diffusion Model (CDM)—were implemented and evaluated. A dataset comprising 503 individual lettuce images with artificially generated random occlusions was used for training and testing. Performance was assessed using pixel-level metrics (PSNR, SSIM) and, more importantly, by evaluating the fresh weight estimation accuracy (R², NRMSE, MAPE) of a pre-trained CNN model (CNN_284) using the inpainted images. Results indicated that AOT-GAN achieved the best overall performance, with an SSIM of 0.9379 and an R² of 0.8480 for weight estimation after inpainting under single-direction occlusion, closely matching the performance using original non-occluded images (R² = 0.8365). In complex multi-direction occlusion scenarios, AOT-GAN demonstrated superior robustness, maintaining an R² of 0.7914 and an MAPE of 12.02% for weight prediction, significantly outperforming the other models. This study demonstrates that advanced inpainting techniques, particularly AOT-GAN, can effectively mitigate the impact of occlusion, enhancing the reliability of vision-based leafy vegetable biomass estimation in practical production.

Reliable scene segmentation is critical for autonomous disinfection robots in complex cattle farming environments, where cluttered backgrounds and severe illumination variation pose challenges to both the segmentation accuracy and real-time inference. However, existing semantic segmentation models often struggle to effectively balance accuracy and inference efficiency. In this study, a lightweight and robust model (CTA-HEFS-TransSeg) was proposed for scene segmentation in complex cattle barns. The CTA-HEFS-TransSeg integrated the Contextual Transformer Attention mechanism (CTA) for enhanced local-global contextual modelling and introduced a novel lightweight Hierarchical Efficient Feature Selection module (HEFS) for efficient multi-scale feature fusion. The results showed that both the CTA and HEFS significantly contributed to the model segmentation performance. When both the CTA and HEFS were jointly integrated, the proposed CTA-HEFS-TransSeg achieved the best segmentation performance with a mean intersection over union (MIoU) of 91.50% and a mean pixel accuracy (MPA) of 95.06%, outperforming the state-of-the-art semantic segmentation models. In addition, the CTA-HEFS-TransSeg maintained real-time performance at 135.3 frames per second (FPS) with only 16.40 M parameters, indicating a favorable balance between segmentation accuracy and computational efficiency. The CTA-HEFS-TransSeg also demonstrated remarkable cross-illumination generalization and multi-scale feature fusion capabilities, robustly generalizing to nighttime cattle barn environments with an MIoU of 77.93% despite being trained exclusively on daytime data. This study indicates that the CTA-HEFS-TransSeg is an efficient and robust tool for visual navigation in complex cattle barn environments.