With the rapid advancement of information technology, three-dimensional (3D) simulation has been increasingly applied in agricultural, showing significant potential for enhancing digital production processes in solar greenhouses. This paper examines the current state of 3D simulation technologies in solar greenhouse production and explores future development trends. It begins by outlining the structural characteristics of solar greenhouses and introducing key enabling technologies for practical application, including computational fluid dynamics (CFD), functional-structural plant modeling (FSPM), and building information modeling (BIM). A comprehensive analysis of application cases is then presented across three critical domains: greenhouse microclimate assessment, crop growth simulation, and structural optimization. These applications cover various stages, from design and planning to crop production and environmental control. Finally, this study identifies current technical challenges and suggests future research directions to advance the use of 3D simulation in smart facility agriculture. Overall, digital twin technology based on realistic canopy structures, combined with AI-driven greenhouse control methods, demonstrates the greatest potential for application. Future studies should prioritize the integration of multi-source heterogeneous data and the improvement of real-time responsiveness in 3D models. By continuously refining data integration and algorithmic models, smart greenhouses can provide more precise and efficient support for sustainable agricultural production.

Forest ecosystems play a pivotal role in maintaining the balance of the global carbon cycle and conserving biodiversity. High-density point clouds derived from unmanned aerial vehicle (UAV) structure from motion (SfM) and multi-view stereo (MVS) technologies offer a cost-effective solution for data acquisition. These technologies have become efficient tools for facilitating precision forest resource management and extracting individual tree structural parameters. However, in complex forest scenarios during the leaf-off season, canopies exhibit unstructured branch network morphologies due to the absence of leaf occlusion, and adjacent crowns are heavily interlaced. Consequently, existing segmentation methods struggle to overcome challenges associated with fuzzy boundaries and instance adhesion. To address these challenges, this study proposes TreeSeg-Net, an end-to-end instance segmentation network designed to precisely separate individual trees directly from raw point clouds. The network incorporates a global context attention module (GCAM) to capture long-range feature dependencies, thereby compensating for the limitations of sparse convolution in perceiving global information. Simultaneously, a spatial proximity weighting module (SPWM) is designed. By introducing geometric center constraints and a distance penalty mechanism, this module effectively mitigates under-segmentation issues caused by the feature similarity of adjacent branches in high-canopy-density environments. Experimental results demonstrate that TreeSeg-Net achieves an average precision (AP) of 97.2% in instance segmentation tasks and a mean intersection over union (mIoU) of 99.7% in semantic segmentation tasks. Compared to mainstream networks, the proposed method exhibits superior segmentation accuracy, providing an efficient and automated technical solution for precise resource inventory in complex forest environments.

Boll-opening concentration is critical for mechanical cotton harvesting, yet it is still assessed mainly by manual records and single time-point indicators that miss temporal dynamics. To bridge the lack of a unified workflow linking vision foundation models, multi-temporal boll-opening monitoring, and harvest decision-making, we developed a cross-scale UAV high-throughput phenotyping framework centered on DINO-BollGX. DINO-BollGX couples a DINO v3 backbone, a RetinaNet detection head, and an adaptive refinement-and-suppression module for robust open-boll detection under complex field conditions. Using multi-temporal UAV imagery collected over two years for 383 cultivars, we reconstructed plot-scale time series of open-boll counts, derived dynamic features describing progression and intensity changes, and proposed a Cotton Boll-Opening Temporal Stability Index (CTSI) to quantify boll-opening rhythm and concentration; CTSI was further integrated with a time-based risk function to generate harvest decision curves. Under unified data and training settings, DINO-BollGX achieved precision = 0.91, F1 = 0.88, and AP@0.50 = 0.80, and provided accurate boll-count estimation (R2 = 0.98; MAE = 3.10), outperforming YOLOv11, YOLOv12, YOLOv13, and RT-DETR. On an independent cross-year dataset acquired at 5 m altitude, it obtained precision = 0.98 and F1 = 0.87. An internal consistency analysis showed that CTSI had the expected negative association with Window_days (r = -0.83) and positive associations with Max_count (r = 0.78) and the boll-opening efficiency index (r = 0.92), reflecting the co-occurrence of temporal compactness and main-phase opening intensity in the cultivar population. CTSI ranged from -2.72 to 4.69 across cultivars, enabling identification of highly synchronized boll-opening. Harvest decision curves indicated that the relative net income index peaked at day 67 after the first observation and a compact optimal harvest window near the end of monitoring; on a fixed harvest date, Kuche 130,292 (CTSI = 4.69) produced 486 open bolls versus 182 for Xinluzao 36 (CTSI = 0.53) and 90 for Andizhan-60 (CTSI = -2.72). Overall, the framework integrates dynamic boll-opening phenotyping with harvest timing optimization, supporting scalable cultivar screening and mechanization-ready deployment, with potential extension to harvest decision scenarios in other crops.

Aboveground biomass (AGB) is a critical indicator for assessing crop growth status and productivity, yet accurately linking fine-scale ground measurements with coarse-resolution satellite imagery remains challenging. Here, we propose an integrated ground-UAV-satellite framework that combines high-resolution UAV observations with an optimized systematic sampling-Global Moran's I (SS-GMI) procedure and a simple allometric growth model. Multi-variety sugar beet cultivated across heterogeneous habitats was used as a case study. Results indicate that a power-law model effectively captures the allometric relationships between AGB, plant height, and the Dreg vegetation index in sugar beet, achieving high accuracy and strong transferability. Incorporating phenological information from Biologische Bundesanstalt, Bundessortenamt und CHemische Industrie (BBCH) codes and a thermal index further enhanced model robustness across independent habitat trials, yielding coefficients of determination (R2) of 0.80 and 0.83. The SS-GMI sampling procedure integrates systematic sampling with Global Moran's I to reduce spatial autocorrelation while ensuring uniform spatial coverage, thereby enabling the acquisition of representative and spatially independent samples from UAV-derived AGB maps. These samples were used to develop satellite-based AGB estimation models for PlanetScope and Sentinel2A imagery, achieving R2 values of 0.83 and 0.73, respectively. This study provides a practical and scalable framework for field-to-satellite AGB upscaling, offering new insights for the scale conversion of multi-source data in agricultural remote sensing.

Soil properties are vital for soil health, fertility, and crop productivity, yet traditional methods like the Partial Least Squares Regression (PLSR) predict individual properties without capturing their interactions or impacts on crop growth. This study presents the Soil Knowledge-Guided Multi-task Transformer (KGMT) model, which integrates deep learning and multi-task learning to simultaneously predict multiple soil properties and crop traits using UAV hyperspectral imagery, eliminating complex manual feature extraction. By embedding soil knowledge, KGMT improves prediction accuracy and generalization over PLSR and baseline multi-task models. Data from two agricultural plots in Bayannur City, Inner Mongolia, China, demonstrated KGMT's superior performance in predicting six soil properties (alkalinity, salinity, organic matter, potassium, sodium, calcium) and two crop traits (leaf area index, aboveground biomass), achieving up to 20% lower normalized root mean square error. For instance, KGMT yielded higher R2 values for alkalinity (0.57) and organic matter (0.57) compared to PLSR (0.46 and 0.40, respectively). The model also enhanced cross-region generalization. Additionally, a soil early warning framework was developed, linking pre-planting soil conditions with crop traits to detect potential growth limitations early, enabling proactive management. This approach integrates spectral data with domain knowledge, offering a robust solution for soil and crop management. It contributes to precision agriculture by optimizing soil conditions and promoting sustainable land use practices, highlighting the value of advanced deep learning in agricultural applications.

The phenotypic traits of beetroot are of great significance for revealing the relationship between their morphology and functional traits, as well as for the breeding high-quality cultivars. The 3D models can provide a more comprehensive and accurate description of beetroot structural features, offering richer phenotypic information compared to traditional 2D imaging methods. This study developed a novel framework that integrates Neural Radiance Fields (NeRF) for high-fidelity 3D reconstruction, multi-dimensional phenotypic traits extraction and an end to end network, named as Point-SugarRoot based on point cloud and deep learning for precise classification. The results indicated that NeRF-derived phenotypic traits strongly correlated with actual measurements, achieving R2 values exceeding 0.92. NeRF-reconstructed point clouds exhibited less noise, with an average C2C distance of 0.11 cm, a SC rate of 94.6%, and a 12.9% reduction in reconstruction time, outperforming the Structure from Motion (SfM)-Multi-View Stereo (MVS) method in precision and efficiency. The Point-SugarRoot network surpassed the baseline PointNet++ model, delivering Precision, Recall, F1 score, and Balanced Accuracy of 91.3%, 90.8%, 91.1%, and 93.0%, respectively, and outperformed phenotype- and 2D image-based classification methods by 13% to 33% across all metrics. This research presents a novel solution for high-fidelity 3D reconstruction and phenotypic analysis of beetroots, while also providing valuable technical insights for cultivar breeding and crop management.

Plant height is a key 3D phenotypic trait for assessing crop growth, biomass accumulation, and lodging resistance. To overcome the practical limitations of conventional plant height measurement methods, this study proposes a novel vision foundation model-based framework named Depth for Plant Height (Depth4PH) for plant height estimation in agricultural scenes using low-cost monocular RGB imaging. As part of our contributions, a synthetic–real coupled multimodal dataset was constructed by integrating Blender virtual agricultural scenes (Blender VAS) with real field images. Building upon the existing Depth Anything V2 foundation model, we developed a novel module called Transfer-based Agricultural Metric Depth Anything V2 (TAM-Depth V2) for absolute metric depth estimation through parameter-efficient fine-tuning, depth decoder reconstruction, and joint loss optimization. Furthermore, we designed a novel multi-source prompt-based segmentation framework, MSP-SAM2, to generate positive and negative prompts for zero-shot crop instance segmentation. Finally, a new inverse physical plant height estimation algorithm, RANSAC-Per, was introduced to estimate plant height by combining truncated percentile statistics with local RANSAC micro-plane fitting, thereby reducing the effects of depth noise and field microtopographic variation. The result showed that TAM-Depth V2 achieved stable absolute depth estimation, with an RMSE of 0.1162 m and an AbsRel of 4.25%. Compared to the original box-prompted SAM 2, MSP-SAM2 achieved a 4.4% improvement in mIoU, reaching 91.6% and a recall of 93.2%. On a 350-plant multi-crop test set, Depth4PH achieved R2 = 0.948, RMSE = 12.23 cm, and MAE = 8.82 cm, and MAPE =10.15%, with crop-specific RMSEs ranging from 3.99 cm (cucumber) to 20.73 cm (maize), significantly outperforming the traditional Global-MinMax baseline (which had an RMSE of 22.62 cm). These results indicate that Depth4PH provides a promising foundational pathway for high-throughput crop phenotyping. With future optimization for edge deployment, it holds significant potential to support high-throughput monitoring in precision agriculture.

High-precision plant phenotyping requires efficient 3D reconstruction with high fidelity, yet existing methods such as MVS and NeRF all have problems of feature dependence and error accumulation during 3D reconstruction, which leads to geometric distortion in reconstruction and restricts the reconstruction efficiency. To address this bottleneck, this study first determined the multi-view image acquisition strategy. Further, based on the self-built multi-view dataset of chili peppers, it proposed an algorithm for efficient and high-fidelity 3D reconstruction of complex plant structures through global adaptive pose optimization and gaussian splash rendering technology, referred to as the GAPose-GS algorithm. Experimental results indicate that the Peak Signal-to-Noise Ratio (PSNR) improves by 52.0 %, 26.4 %, and 4.2 % compared to NeRF, Instant-NGP, and 3D Gaussian Splatting respectively. Additionally, the Structural Similarity Index Measure (SSIM) increases by 22.9 %, 12.8 %, and 4.3 % respectively over above methods. The point cloud data reconstructed based on this algorithm also has advantages in the measurement of phenotypic parameters. Compared with the actual measured values, the R2 of the phenotypic parameters such as pepper plant height, canopy width, and leafstalk angle obtained in this study are 0.997, 0.954 and 0.978 respectively, and the RMSE are 0.236 cm, 1.082 cm and 2.344 degrees respectively, and the MAE are 0.209 cm, 0.880 cm and 1.965 degrees respectively. The accuracy was significantly better than that of the existing phenotypic calculation methods. Verification across different growth stages of wheat and maize was performed universally, with all errors remaining below 1.1 %, providing new ideas and technologies for high-precision, low-cost, and high-throughput crop phenotypic research.

Canopy chlorophyll density (CCD) is a vital indicator for evaluating maize nitrogen status and guiding precision fertilization. Traditionally, CCD is derived from destructive field sampling followed by laboratory analysis. Recent advances in unmanned aerial vehicle (UAV)-based hyperspectral imaging offer a rapid, non-destructive alternative for real-time CCD estimation. Conventionally, CCD is calculated as the product of leaf chlorophyll content and leaf area index (LAI), yet LAI is sensitive to varietal differences and may introduce uncertainty. To address this limitation, an improved canopy chlorophyll density (ICCD) metric was proposed by replacing LAI with canopy coverage (CC). The ICCD and CCD were compared in terms of spectral feature selection and estimation accuracy. In addition, the influence of spectral bandwidth on UAV-based hyperspectral estimation of maize CCD before and after the improvement was evaluated. The results indicated that the ICCD estimation model constructed using UAV hyperspectral data consistently demonstrated higher accuracy compared to that of the CCD model. On the testing set, the ICCD estimation achieved R2 values of 0.7903-0.8161, RPD values of 2.1573-2.3248, and RMSE values of 5.9574-6.4200. In contrast, the CCD estimation yielded R2 values of 0.7331-0.7954, RPD values of 1.9421-2.2197, and RMSE values of 48.4747-55.4027. Additionally, increasing the spectral bandwidth from 4 nm to 24 nm resulted in a stepwise decline in R2 and RPD values, along with a corresponding increase in RMSE for the ICCD model. In summary, the "CC instead of LAI" strategy enhanced the robustness and accuracy of CCD estimation based on UAV hyperspectral imagery. Moreover, the analysis of bandwidth effects can provide valuable guidance for the design and optimization of future multispectral sensors in the field of precision agriculture.

Non-destructive real-time monitoring of stem-leaf fresh biomass during the vegetative and reproductive stages in greenhouse tomatoes is essential for optimizing fruit management. Traditional RGB-based methods frequently encounter signal saturation due to canopy closure and lack critical plant physiological information. To address these limitations, an integrated RGB, multispectral, and thermal infrared (RGB-MS-TIR) monitoring system was developed presenting a "Volume-Density-Vitality" analytical framework. First, the ZoeDepth and EfficientSAM algorithms were utilized to extract plant height and canopy cover with high precision, establishing a structural 'volume' skeleton (R-2 = 0.90, RMSE = 12.43 cm). Second, multispectral indices were employed to quantify chlorophyll 'density', while thermal infrared features were used to characterize physiological 'vitality'. The results demonstrated that when canopy cover exceeded 75 % (the saturation zone), the canopy-air temperature difference (Delta Tc-a) exhibited a significant negative correlation with biomass (P < 0.001), effectively mitigating geometric signal saturation through the transpiration cooling effect. Ultimately, the Random Forest (RF) model integrating RGB, MS, and TIR features achieved the superior performance in estimating stem-leaf fresh biomass, with a coefficient of determination (R-2) of 0.96 and an RMSE of 46.28 g plant(-)& sup1;. This research highlights that multimodal fusion significantly enhances estimation accuracy under complex canopy conditions, providing a robust solution for intelligent greenhouse crop management.

Light is essential for photosynthesis and directly influences crop yield. During winter and spring, limited natural light makes well-managed supplemental lighting crucial for greenhouse production. Traditional lighting design methods, which rely on manual measurements, are inefficient for optimizing light distribution and energy use. This study proposes a 3D simulation framework to optimize supplemental lighting in greenhouses. The virtual model incorporates the spectral power distribution (SPD) and propagation characteristics of light-emitting diode (LED) modules, the optical properties of greenhouse materials, and the greenhouse's geometric structure to simulate artificial light environments. Validation of the model demonstrated high accuracy, with an R2 of 0.982 and a RMSE of 14.38 mu mol center dot m-2 center dot s-1. Based on simulation outputs, the spatial layout of supplemental lighting modules was determined, and the hourly light integral (HLI) was used as a control variable to develop a timesegmented lighting strategy. For this study, the production performance of tomato was evaluated under four lighting treatments: HLI-driven fixed supplementary lighting (HFS), HLI-driven mobile supplementary lighting (HMS), nighttime timed supplementary lighting (TS), and only natural light (CK). The optimal lighting configuration was achieved when fixtures were positioned 1.7 m above the planting troughs. Tomato yield per plant under the HFS treatment increased by 25.2% compared to CK and by 21.6% compared to TS. While HMS showed higher energy-use efficiency and quantum yield, its yield improvement was relatively modest. Overall, HFS enhanced light energy-use efficiency and quantum yield by 5.5% and 55.3%, respectively, compared to TS. This study provides a practical decision-support tool for greenhouse lighting management, enabling data-driven optimization of light distribution and energy use. The proposed 3D modeling framework not only improves light-thermal synergy but also offers strong scalability for different greenhouse structures and crops. By integrating physical modeling and intelligent control, it contributes to the development of sustainable and smart agricultural production systems.

The estimation of lettuce fresh weight is critical for assessing growth status and optimizing cultivation. Traditional methods are often inefficient, error-prone, and costly. Computer vision offers opportunities for image-based non-destructive fresh weight estimation. This paper introduces MIFFNet, an end-to-end network integrating RGB images, visible light vegetation indices, geometric features, and color features for lettuce fresh weight estimation. This model employs the Inception structure as the multi-scale feature extraction (MSFE) block, alternating with the proposed multidimensional image feature fusion (MIFF) module to form the network backbone. This design enhances the model’s ability to capture multiscale features while thoroughly integrating multidimensional image features. Comparative experiments were conducted with 10 competitors, including classical convolutional neural networks, and existing lettuce fresh weight estimation models, across three lettuce datasets. Experimental results demonstrated that MIFFNet outperforms others across all three datasets. On the self-built dataset, it achieved an R2 of 0.929, with RMSE and MAE values of 28.544g and 14.446g, respectively. On two public datasets, the R2 values reached 0.94 and 0.943, with lower RMSE and MAE than competitors. Furthermore, MIFFNet exhibited significant advantages in terms of model complexity and parameter efficiency. These results highlight MIFFNet’s superior capability of accurate and efficient lettuce fresh weight estimation.

Precision agriculture requires accurate yet efficient crop organ detection methods that can be readily deployed under field conditions with limited storage and computing resources. This study therefore examined whether the independently acquired multispectral red single-band image (MS-R) can provide a more effective alternative to conventional RGB images for cotton flower detection. We compared its performance with representative YOLO-series detectors, two mainstream non-YOLO detectors, and progressive ablation variants using MS-R, RGB, and RGB-R datasets. The evaluation metrics included Precision, Recall, mAP, storage efficiency, and computational cost. Additionally, we investigated the effect of MS-R image re-encoding with image-quality parameters down to N = 25 on detection performance and storage efficiency. Across the nine compared detector variants, paired significance tests confirmed that MS-R achieved significantly higher Precision, Recall, and mAP than RGB and RGB-R. The exact P values ranged from 0.008 to 0.012. Specifically, the MS-R data achieved a Precision of 0.911, compared with 0.862 for RGB, and also showed higher Recall (0.917 vs. 0.830) and mAP (0.953 vs. 0.881). Moreover, the average size of MS-R images is 55

High-throughput field phenotyping bridges genotype, environment, and phenotypic performance. Conventional plot-level approaches relying on manual surveys are labor-intensive, and error-prone and fail to capture variability among individual plants, limiting seed cotton yield estimation and genotype screening under natural conditions. To address these limitations, a complex framework was developed, integrating single-plant instance segmentation, multi-trait inversion, plot-level stability characterization, and yield estimation. The enhanced vision-model framework, TopoRefineSAM, combines YOLOv12 detection with SAM2 segmentation and incorporates adaptive enhancement and topological refinement modules, enabling efficient, robust, and cost-effective single-plant identification under weak annotation. Based on this segmentation, multi-source UAV imagery (RGB, multispectral, thermal infrared, and DSM) was used to build ensemble learning models for inversion of physiological and biomass traits. A Stability Index Group (SIG) translates inter-plant variability into plot-level stability features, improving interpretability and consistency in yield estimation and cultivar screening. Results demonstrated that TopoRefineSAM achieved high segmentation accuracy for single-plant extraction under complex field conditions. In multi-trait inversion, Gradient Boosting Decision Trees (GBDT) achieved the highest performance. Our results demonstrated strong consistency between multimodal features and measured traits. In yield estimation, incorporating the SIG substantially improved predictive performance across growth stages. In cultivar screening, the method achieved high agreement with field measurements, showing robust identification of top-performing cultivars. Collectively, the findings establish a scalable, cost-effective, and high-accuracy framework for field-based phenotypic analysis and yield estimation, providing both methodological innovations and practical support for precision breeding and large-scale crop improvement.

Rice disease identification is a critical technique for ensuring yield and quality in precision agriculture. However, complex field backgrounds, subtle lesion features, and similar symptomatic manifestations have led to low accuracy and poor robustness in traditional classification methods. To address these issues, this study proposes an improved ConvNeXt network model (Improve‑ConvNeXt) for the fine classification of rice diseases under field conditions. A high‑quality dataset containing six categories (healthy rice, rice blast, brown spot, bacterial leaf blight, bacterial leaf streak, and bacterial grain rot) was constructed from field images and public datasets, with a total of 5,663 samples. Using ConvNeXt‑Tiny as the backbone, the model integrates a Hybrid Attention Transformer (HAT) to enhance the perception of lesion regions and key channels, and introduces Spatial and Channel Reconstruction Convolution (SCConv) to reduce feature redundancy and strengthen effective information expression. Experiments show that the Improve‑ConvNeXt model achieves 96.27% accuracy on the test set, which is 4.85% higher than that of the original ConvNeXt and significantly outperforms ResNet and DenseNet. The precision, recall, and F1‑score reach 95.84%, 96.11%, and 95.95%, respectively. Confusion matrix and Grad‑CAM visualization prove that the model can accurately focus on lesion areas and effectively distinguish similar diseases. This method provides high precision and strong generalization for rice disease identification in complex field environments, and offers a reliable technical reference for intelligent monitoring and precise management of rice fields.

[Background and objective] Maize/soybean intercropping performance is strongly influenced by three-dimensional (3D) canopy architecture, which regulates light interception and crop complementarity. This study aimed to reconstruct field maize/soybean canopies from unmanned aerial vehicle (UAV) RGB imagery and quantify how planting pattern and density affect canopy light distribution, yield, and land equivalent ratio (LER).[Method] Field experiments were conducted in northeast China using three row configurations (2:2, 2:4, and 4:4 maize/soybean strips) and three maize planting densities (6.4 × 104, 9.6 × 104, and 11.2 × 104 plants ha⁻1). Canopy structure was reconstructed using UAV RGB cross-surround imaging, structure-from-motion and multi-view stereo (SfM-MVS), and point-cloud processing. Reconstruction accuracy was evaluated using plant height, leaf length, and leaf width. The Helios light model was used to simulate canopy light interception and distribution.[Results] Plant height estimation achieved R2 values of 0.92 for maize and 0.72 for soybean, while leaf length and width estimates achieved R2 ≥ 0.88 and relative root mean square errors of 5–10 %. Helios simulations showed treatment-specific differences in canopy light distribution. At 102 days after emergence, maize light interception rate in medium-density 2:2 and 2:4 treatments exceeded 600 W m⁻2 during peak sunlight hours, whereas high-density treatments increased soybean shading. Medium-density treatments showed a favorable numerical balance between light capture and crop complementarity, with LER values of 1.06 ± 0.16 for 2:2M and 1.00 ± 0.04 for 4:4M; however, total LER was not significantly affected by planting density.[Conclusion] UAV RGB cross-surround imaging combined with SfM-MVS reconstruction and light-interception modeling provides an effective workflow for linking 3D canopy architecture, light acquisition, and yield formation. Medium-density configurations showed a favorable numerical balance between canopy light capture and interspecific complementarity.

Accurate prediction of temperature and humidity fields in solar greenhouses is essential for precise environmental control and disease risk early warning. However, the microclimate exhibits significant spatiotemporal heterogeneity due to the asymmetrical envelope structure and the development of the crop canopy in solar greenhouses. Consequently, conventional single-point monitoring or purely time-series models struggle to accurately characterize its three-dimensional spatial distribution. To address this challenge, this study proposed a multi-sensor deep learning framework that integrated three-dimensional spatial coordinate embedding and sensor-level gated attention mechanisms. A CNN-BiLSTM-SA model was developed, which extracted local dynamic features via multi-scale convolution, captured long-term dependencies of environmental parameters using BiLSTM, and dynamically assigned weights to different sensors through a sensor-level gated attention mechanism according to environmental states. This enabled multi-step prediction and continuous three-dimensional reconstruction of temperature and humidity fields. The model was validated using multi-source sensor monitoring data collected throughout the entire growth cycle of cucumbers in a typical solar greenhouse. The results indicated that the proposed model achieves MAE, RMSE, and R² of 0.88°C, 1.236°C, and 0.92 for temperature, and 2.619%, 3.735%, and 0.93 for relative humidity. Relative to CNN-BiLSTM, sensor-level gated attention reduced temperature and humidity RMSE by 26.6% and 37.9%. Two-dimensional slices, three-dimensional field reconstructions, and time-series analysis at key nodes further demonstrated the model’s ability to effectively recover the dominant spatial structures and dynamic evolution patterns of the temperature and humidity fields. This study has achieved a precise characterization of the temperature and humidity fields in solar greenhouses, providing a high-accuracy data foundation and methodological support for zoned precise regulation and disease risk warning in solar greenhouses.

Background and aim Maintaining an appropriate nitrogen (N) status is crucial for achieving desirable grain quality in barley. N fertilization is typically applied at sowing, leaving limited scope for in-season adjustment. Timely monitoring is essential for grain N concentration prediction and quality assessment. This study aimed to develop an artificial intelligence (AI) driven framework to estimate N concentration, N uptake, N nutrition index (NNI), and dry matter (DM) using unmanned aerial vehicle (UAV)-based multispectral imagery collected at three key growth stages (jointing, DC30; heading, DC55; and flowering, DC65). Methods Two machine learning algorithms, random forest (RF) and Keras-based feedforward neural network (KFNN), were used to estimate N-related indicators. Multiple vegetation indices (VIs) were extracted from multispectral data. Feature importance ranking and non-dominated sorting genetic algorithm (NSGA) were applied to optimize input variables for improving model performance and robustness. Results The results showed that N-related indicators were better estimated at DC55 and DC65 stage, while DM were well predicted at DC30 stage. KFNN consistently outperformed RF in modeling complex traits such as N uptake and DM, with D-index improvements exceeding 10%. NSGA-optimized feature sets outperformed traditional importance ranking by improving both model stability and predictive accuracy. The top 3 VIs combinations achieved a favorable balance between accuracy and redundancy. All models achieved D-index values above 0.75 across N-related indicators and DM. Conclusion This study demonstrates that combining growth-stage-specific modeling with AI-based feature optimization provides a scalable and cost-effective approach for real-time N monitoring, supporting precision N management in spring barley.