Open research questions in Smart Agriculture and AI

What the literature leaves open

• By contrast, PAC-Bayesian analysis provides a theoreti- cally founded means to quantify uncertainty related to model generalization in limited data settings.

  PDRPFRS: a hierarchically integrated ensemble CNN–Bayesian fertilizer recommendation framework with statistical field validation for precision phytodiagnostics · 2026 · DOI

• The lack of standardized, interoperable, and representative datasets limits reproducibility, transferability, and scalability of AI systems, thereby constraining their deployment in operational agricultural applications.

  A multiregional image-text dataset and benchmark for vision-language modeling of plant diseases · 2026 · DOI

• Most cited studies are proof-of-concept rather than large-scale field implementations, and the generalizability of AI models across different epidemiological settings remains to be 250 Proceedings of ICEGEE 2026 Symposium: AI-Based Medicine and Biological Data Analysis DOI: 10.

  Application and Prospects of Artificial Intelligence in Schistosomiasis Prevention and Control · 2026 · DOI

• application on the percentage of product based International Research Journal on Advanced Engineering Hub (IRJAEH) 3129 International Research Journal on Advanced Engineering Hub (IRJAEH) e ISSN: 2584-2137 Vol. 04 Issue: 03 March 2026 Page No: 3128-3133 https://irjaeh.com https://doi.org/10.47392/IRJAEH.2026.0397 classification/severity level as well as the disease name; targeted product applications rather than blanket spray applications.

  An AI-Driven End-to-End Agricultural Guidance System with Multilingual and Voice Support · 2026 · DOI

• A key observation is that traditional tabular features, weight and shape index, long-standing tabular bench- marks in egg-grading standards [12, 13, 68], are insufficient for reliable classification, as demonstrated by their poor predictive performance in this study.

  ELMF4EggQ: ensemble learning with multimodal feature fusion for non-destructive egg quality assessment · 2026 · DOI

• Survey-only study without experimental implementation; scarcity and bias in agricultural datasets; poor cross-crop and seasonal generalization; computational and energy limitations of edge hardware; lack of standardized benchmarks; socio-economic and adoption barriers not empirically evaluated.

  Edge-AI for Real-Time Agricultural Monitoring: A Review and Conceptual Framework for Organic Farming · 2026 · DOI

• Future work will focus on improving robustness at early growth stages (especially BBCH 15), investigating adaptive or learned vegetation descriptors, validating the method on larger multi-site and multi-season datasets, and optimizing the framework for real-time or near real-time embedded deployment in UAV- or robot-assisted precision agriculture systems.

  Annotation-efficient weed mapping in sorghum fields using two-stage U-Net crop segmentation and HSV greenness analysis · 2026 · DOI

• Limitations of Low-Cost Sensors Although the selected sensors provided acceptable performance for trend-based agricultural monitoring, their should be considered when interpreting the results. The DHT22 sensor is suitable for general temperature and humidity monitoring, but its response may be slower under rapidly changing humidity conditions. Its readings may also be affected by condensation, dust, and long-term exposure to outdoor environments. Therefore, periodic checking or simple recalibration is recommended when the sensor is used for extended field deployment. The BMP180 pressure sensor showed stable shortterm readings in this study; however, pressure measurements may still be affected by altitude differences, local microclimate variation, enclosure design, and temperature compensation. For this reason, the pressure values should be interpreted mainly as environmental trend indicators rather than as highly precise meteorological measurements. the The MQ-2 gas sensor has more obvious limitations compared with temperature, humidity, and pressure sensors. Since it is an analog low-cost gas sensor, its output can be influenced by temperature, humidity, warm-up time, sensor aging, and crosssensitivity to different gases. In this study, the MQ-2 was therefore used only to detect relative changes between normal and abnormal conditions, rather than to estimate absolute gas concentration. This interpretation is consistent with the practical purpose of the proposed system, which focuses on low-cost environmental monitoring and early indication of abnormal field conditions. Overall, the use of low-cost sensors makes the proposed system more accessible for agricultural deployment, but it also introduces a trade-off between affordability and measurement precision. The results should therefore be interpreted as suitable for practical monitoring, trend observation, and field-level decision support, rather than as laboratory-grade measurements. 3.4.3 Impact of Missing Data and Packet Loss To evaluate the impact of communication loss, periods of packet loss were simulated, and the resulting data availability was analyzed. Table 11 indicates that even under moderate packet loss conditions, the overall data trends remain interpretable. This robustness is attributed to the periodic sensing strategy and time-series analysis at the backend, which compensates for occasional missing data. Table 11.

  An Integrated LoRa-Enabled IoT Sensing System for Precision Agriculture: Design Algorithm and Practical Evaluation · 2026 · DOI

• ARTICLE IN PRESS ARTICLE IN PRESS ACCEPTED MANUSCRIPT Although EfficientNetV2 provides strong accuracy and efficiency, training the Medium (M) and Large (L) variants on a CPU remains time- consuming. This limitation makes it difficult to explore deeper models fully and increases the overall training time for experiments. This study utilized a single publicly available dataset with only three disease classes (Healthy, Rust, Spot), which may not fully represent the complexity of real-world agricultural scenarios where multiple diseases and mixed infections may occur simultaneously. The original dataset is relatively small for training deep learning models. While augmentation expanded the dataset images, synthetic bias may be introduced if augmented images overrepresent certain features or artifacts. Future work will focus on aggregating multiple publicly available mulberry leaf datasets from different online repositories to create a larger, more diverse, and cross-validated corpus. This multi-dataset approach would enhance model generalizability across different geographic regions, imaging conditions, and disease presentations. Due to computational constraints (CPU-only training with limited RAM), k-fold cross-validation was not performed in this study. Instead, a fixed train-test split (80/20) was used for evaluation. While this approach is common in similar agricultural AI studies, cross-validation would provide more robust performance estimates with standard deviations. This study compared only EfficientNetV2 variants (S, M, and L) due to computational constraints. Future work will include comprehensive benchmarking with standard baseline architectures such as ResNet50, MobileNetV2, DenseNet121, and Vision Transformers to further validate the proposed approach. Although the proposed model achieves high classification accuracy, it currently operates as a black-box system. Future work will incorporate more advanced explainable AI techniques such as Grad- CAM, LIME, and SHAP to provide visual interpretations of model predictions, thereby in real-world agricultural applications. improving transparency and trust Contributions: All authors contributed to the study’s conception and design. Material preparation, data collection and analysis were performed by BKN, KKC, RRK and SL. The first draft of the manuscript was written by BKN, RRK and KKC commented on its improvement. Reviewing is done by PKA, SJ, SM, STH. All authors read and approved the final manuscript.

  AI-driven mulberry leaf disease detection for sustainable and resilient sericulture systems · 2026 · DOI

• Future research should focus on open orchard environments and develop data-efficient, interpretable, low-power, and continuously updatable edge-intelligent recognition systems, thereby advancing precision agriculture and smart orchards.

  Deep Learning-Based Fruit Tree Pest and Disease Recognition Technology: Model Evolution, Challenges, and Edge Intelligence Deployment · 2026 · DOI

• Rice production in the Lao People’s Democratic Republic (Lao PDR) is fundamental to national food security, yet its productivity and environmental impacts remain poorly quantified due to limited monitoring and data availability.

  A hybrid machine-learning and process-based regional crop modelling framework to improve rice production strategies under data constraints · 2026 · DOI

• Furthermore, the Future work may focus on proposed system shows potential to expanding the dataset with additional support agricultural activities, particularly rhizome spice classes and more diverse in assisting users in identifying rhizome image conditions, as well as investigating spices more efficiently and reducing other lightweight deep learning misidentification.

  COMPARATIVE PERFORMANCE AND GENERALIZATION ANALYSIS OF MOBILENETV1 AND MOBILENETV2 FOR RHIZOME SPICE CLASSIFICATION · 2026 · DOI

• The proposed dataset includes images collected from different agricultural regions and environmental conditions. Some limita- tions must be considered when evaluating the results of this study. The dataset was created from data obtained from three different provinces of Türkiye under varying light conditions, soil structures, weather conditions, and image acquisition angles. The aim was to increase environmental diversity and improve the robustness of the models in real agricultural conditions. The dataset was collected using drones and phones. Although collected from three different provinces and different fields in Türkiye, the dataset still represents a limited geographical region and climate profile. A significant limitation concerns the generalization ability of the models across different seasons and plant development stages. The dataset generally includes visually distinguishable stages of weeds. Model performance may vary in different phenological stages, seasonal conditions, or agricultural practices. Different camera systems, sensor qualities, and image resolutions can also affect model performance in real field applications. The proposed benchmark demonstrated high detection perfor- mance under real field conditions. Validation supporting the results using independent datasets from different countries or agricultural environments could not be performed. Therefore, future studies require a broader review of the benchmark to better assess the model’s transferability and generalization capacity. The practical use of deep learning-based weed detection systems can also present various operational challenges. Changes in lighting, shadows, camera lens contamination, overlapping weeds, motion- induced image distortions, wind, and hardware limitations associ- ated with embedded systems can negatively impact real-time detection performance. Future studies are expected to focus not only on expanding the dataset under different environmental conditions and seasons, but also on developing lightweight and efficient application strategies for drone-based and autonomous precision agriculture systems.

  TomatoweedDet: a real-field multi-class weed detection dataset and YOLO benchmark for tomato production systems · 2026 · DOI

• towards farming, to be more precise. The system might have connected to drones and autonomous vehicles and been adapted to work on dissimilar varieties of crops and agricultural contexts including small farms and large business ventures. The IoT required an upgrade of its infrastructure to safeguard the information of consumers.

  Optimizing Crop Yield and Monitoring Leaves with an Intelligent Internet of Things · 2026 · DOI

• Despite the promising results, several limitations must be acknowledged. First, the dataset used in this study consists of high-quality, single-object images, which do not fully capture the complexity of offshore environments, including motion blur, occlusion, multi-object scenes, and adverse weather conditions. This introduces a domain shift between training data and real-world deployment scenarios. Second, the current implementation focuses on image classification rather than full object detection and tracking. In practical applications, detection and localization would be required as a preceding step before classification. Third, the SCADA integration has been validated within a simulation-based environment. The present study should be interpreted as a simulation-based proof-of-concept framework rather than a field-deployed offshore turbine-control system. The SCADA layer was used for monitoring, communication emulation, and evaluation of turbine-response logic under controlled software conditions. Consequently, the reported latency and response metrics represent feasibility indicators rather than field-validated operational performance a simulation-based. Fourth, the proposed risk-zone mapping is based on simplified geometric assumptions, without direct depth estimation from vision sensors. Future work will explore stereo vision, LiDAR, or radar fusion techniques to enhance spatial accuracy. Future research will focus on deploying the framework using real offshore datasets, integrating object detection models such as YOLO-based architectures, and validating SCADA integration under real operational conditions. Additionally, performance under domain shift conditions may introduce increased false positive and false negative rates, which must be evaluated using real offshore datasets.

  An intelligent SCADA-integrated deep learning framework for bird-safe offshore wind farm operation · 2026 · DOI

• for pest diagnosis and management. Representative Android-based mobile applications used for pest management in various countries are presented in Table 6 (83-92). Mobile applications integrated with AI-based chatbots and cloud computing platforms are emerging as highly effective ICT tools for pest management advisories. They enable the real-time collection, integration and analysis of available data. AI chatbots provide interactive, personalised recommendations, while cloud platforms support large-scale data processing and continuous Table 5. Some of the important SMS and voice call providers for agriculture/pest related inforation in india S.No.

  Evolving pest management paradigms through information and communication technologies · 2026 · DOI

• Addressing the constraints of ICT-based pest management calls for a comprehensive and vision-oriented approach. Rural digital infrastructure investment is essential, particularly in enhancing Internet connectivity and electricity supply. Access to ICT tools must be provided on an equal basis. Developing user-friendly applications in local languages, with voice-based interfaces and intuitive designs, will further enhance accessibility for elderly and low-literacy communities. Advances in AI and machine learning should be supported by the development of large, high-quality, open-source pest image datasets that represent diverse crops and agroecological regions. This will raise the accuracy of computerised pest identification systems in various farming systems. Sensor calibration and farmers training are key to addressing data quality challenges in pest monitoring. Additionally, predictive modelling systems must be continuously refined to incorporate evolving pest dynamics and climate variability. Stronger collaboration among ICT developers, extension agencies, research institutions and private-sector partners is essential for overcoming current limitations. Integrated digital platforms that consolidate pest advisories from multiple sources can https://plantsciencetoday.online reduce contradictory recommendations and promote consistent, science-based guidance. Ensuring the timely dissemination of ICT innovations to farmers will further support the adoption of sustainable, ICT-enabled pest management practices.

  Evolving pest management paradigms through information and communication technologies · 2026 · DOI

• Future work will focus on expanding the dataset to include a wider variety of leaf disease stages and environmental conditions, integrating deep learning architectures such as Convolutional Neural Networks (CNNs) to improve detection robustness and accuracy, and deploying the system as a lightweight mobile or IoT-based application to enable real-time, field-level leaf disease diagnosis. The findings of this study confirm that image processing and machine learning techniques offer a viable and effective alternative to conventional manual leaf inspection methods, which are time- field varying conditions, Despite its promising performance, the system has certain limitations that warrant further investigation.

  Image Processing-Based Detection of Pomegranate Leaf Diseases Using K-Means Clustering and SVM · 2026 · DOI

• This study proposed CFPR-YOLO, a lightweight framework for chili flower detection and pose-aware perception under complex agricultural conditions. The method was designed as an edge- deployable visual perception module for intelligent pollination systems and was evaluated on the constructed dataset, deployed hardware platform, and controlled simulation conditions. By integrating EfficientFormerV2, the C3k2_EMA module, PSConv, and a lightweight attention mechanism, the framework improved the recognition of small, densely distributed, and partially occluded floral targets, while maintaining relatively low computa- tional complexity on the evaluated hardware platform. Experimental results indicate that CFPR-YOLO achieved strong performance on the self-constructed dataset (mAP@50 of 92.1%) and showed improvements over baseline and mainstream models under the same experimental settings. The model also demon- strated stable performance under varying lighting and occlusion conditions within the collected dataset. Deployment on the evalu- ated edge platform verified its real-time inference capability under the tested configuration. Furthermore, under controlled simulated pollination conditions, the system achieved a pollination success rate of 90.0% for upward-facing flowers. These results suggest the effectiveness of the proposed framework within the tested dataset, hardware environment, and controlled scenarios. The contribution of this work lies in developing a perception- to-action pipeline tailored to agricultural engineering informatics, providing reliable visual inputs for robotic pollination, task plan- ning, and precision operations under the tested conditions. However, the findings are limited to the specific dataset, hardware setup, and simulation environment used in this study. In addition, data augmentation was performed prior to dataset partitioning during the current experimental setup. Although repeated experiments under different random seeds yielded rela- tively consistent performance trends, augmented samples derived from the same original image may still introduce potential data correlation across subsets. Therefore, the reported results should be interpreted within the scope of the current dataset and experimental protocol. Future studies will adopt augmentation only on the training subset to further improve methodological rigor and eval- uation reliability. Several limitations remain. The framework shows performance degradation in challenging pose conditions (e.g., backward-facing flowers) within the collected dataset. False positives still occur under strong illumination or structurally similar backgrounds. Additionally, the discrete pose representation restricts fine-grained spatial modeling. The system also relies solely on RGB imagery, which limits its capability to capture complex structural informa- tion under field conditions. Future work will focus on improving robustness and applica- bility under broader conditions. This includes exploring higher- resolution or continuous pose representations, integrating multi- modal sensing data (e.g., depth and spectral information), and further validating the system across more diverse datasets, hardware platforms, and real-world agricultural environments.

  CFPR-YOLO: chili flower pose estimation for robotic pollination in unstructured environments · 2026 · DOI

• ARTICLE IN PRESS ARTICLE IN PRESS ACCEPTED MANUSCRIPT 2024 – CNN-based plant disease recognition with dataset analysis 2016 – Early CNN adoption for plant disease identification 2023 – Deep CNNs with robustness analysis – CNN-based 2019 classification with handcrafted preprocessing 2025 – Advanced deep learning architectures for plant pathology 2020 – Super-resolution (GAN-based) for plant disease images 2023 – Transfer learning with EfficientNet variants…

  Deep convolutional models for robust multi-crop disease recognition in real-world conditions · 2026 · DOI

• To improve reliability, future work will focus on integrating real-time sensor data, satellite imagery, and geo-spatial mapping to enhance feature richness. To address this, future work will focus on enhancing the feature set by integrating real-time weather data, soil sensor inputs, and farmer-specific historical preferences.

  Smart Agriculture: Leveraging Machine Learning for Crop Recommendation, Fertilizer Optimization, and Yield Prediction · 2026 · DOI

• Future research will focus on building a diverse and sufficiently large agricultural disease dataset to enhance the gen- eralization ability of the proposed model. First, due to the limited coverage of existing data- sets, the generalization ability of the proposed model under unseen diseases, different tomato growth stages, and acqui- sition conditions from different regions remains to be further validated.

  TDD-YOLO: A novel model for precise detection of tomato diseases · 2026 · DOI

• The AI-Driven Crop Disease Prediction System bridges conventional agricultural practices with advanced automation. By combining computer vision, deep learning, and environmental data analysis, the system provides farmers with an efficient, accurate, and accessible solution for early disease identification. It reduces dependency on manual expert inspection, supports localized decision-making, and promotes sustainable crop management through timely intervention and prevention strategies. Future work will focus on the following enhancements: • Blockchain-based record storage to ensure transparency and traceability of disease data. • Federated AI learning models for privacy-preserving and region-specific training. • Integration of conversational chatbots to assist farmers with instant, context-aware recommendations. • Predictive modeling for disease outbreak forecasting using weather and soil parameters. • IoT-based real-time monitoring for automated image and sensor data collection. • Cloud-based synchronization for scalable deployment across multiple agricultural zones. © Author(s). This work is peer-reviewed, openly published, and permanently archived This article is openly accessible and reusable with proper attribution.

  AI-Driven Crop Disease Prediction System · 2026 · DOI

• Despite its effectiveness, the proposed system has certain limitations: • Dataset Dependency: model performance depends heavily on the quality and diversity of the training dataset. • Sensitivity to image quality: poor lighting, blur, or noise in image can affect prediction accuracy.

  Plant Leaf Disease Classification Using Convolutional Neural Networks (CNN) · 2026 · DOI

• Future improvement can enhance the performance and applicability of the proposed system: • Integration with Advanced Architecture: Incorporating deeper models such as ResNet can improve feature extraction and classification accuracy.

  Plant Leaf Disease Classification Using Convolutional Neural Networks (CNN) · 2026 · DOI

Most-cited papers in Smart Agriculture and AI

• Agricultural object detection with You Only Look Once (YOLO) Algorithm: A bibliometric and systematic literature review · Computers and Electronics in Agriculture · 2024 · 344 citations
• Plant disease detection and classification techniques: a comparative study of the performances · Journal Of Big Data · 2024 · 306 citations
• Revolutionizing agriculture with artificial intelligence: plant disease detection methods, applications, and their limitations · Frontiers in Plant Science · 2024 · 272 citations
• Integrating artificial intelligence and Internet of Things (IoT) for enhanced crop monitoring and management in precision agriculture · Sensors International · 2024 · 263 citations
• Towards Precision Agriculture: IoT-Enabled Intelligent Irrigation Systems Using Deep Learning Neural Network · IEEE Sensors Journal · 2021 · 239 citations
• Crop pest identification using deep network based extracted features and MobileENet in smart agriculture · Land Degradation and Development · 2024 · 229 citations
• Transformative Technologies in Digital Agriculture: Leveraging Internet of Things, Remote Sensing, and Artificial Intelligence for Smart Crop Management · Journal of Sensor and Actuator Networks · 2024 · 204 citations
• Sustainable AI-based production agriculture: Exploring AI applications and implications in agricultural practices · Smart Agricultural Technology · 2024 · 180 citations
• Comparative performance of YOLOv8, YOLOv9, YOLOv10, YOLOv11 and Faster R-CNN models for detection of multiple weed species · Smart Agricultural Technology · 2024 · 179 citations
• Enhancing crop recommendation systems with explainable artificial intelligence: a study on agricultural decision-making · Neural Computing and Applications · 2024 · 177 citations

Most recent work

• Optimizing Agricultural Disease Detection with Hybrid Weighted Particle Swarm Optimization-based Faster Mask RCNN for Secure Cloud Analysis and Steady Productivity · International Journal of Computational Intelligence Systems · 2026
• PlantPathNet: a novel network based on cross-layer feature integration for plant disease classification · Frontiers in Plant Science · 2026
• Enhanced Lemon Leaf Disease Diagnosis via Gray Wolf Optimization and Machine Learning Techniques · Applied Fruit Science · 2026
• Deep learning techniques for early detection and classification of leaf diseases in crops · Frontiers in Plant Science · 2026
• Crop robots as potential enablers of economical and biodiversity-smart small-scale farming · Precision Agriculture · 2026
• Classification of Rice Leaf Diseases Based on Lightweight Deep Learning Model · ICCK Transactions on Emerging Topics in Artificial Intelligence · 2026
• Robust Crop Planning Under Uncertainty: Aligning Economic Optimality with Agronomic Sustainability · Proceedings of the International Conference on Automated Planning and Scheduling · 2026
• A Lightweight and High-Precision Citrus Detection Model for Unstructured Orchard Environments · Horticulturae · 2026
• Research on Innovative Applications of Generative Artificial Intelligence in Agricultural Informatization · Digital Intelligence in Agriculture · 2026
• Achieving Near-Perfect Performance: A Neuro-Fuzzy-Enhanced Meta-Learning Model for Automated Severity-Graded Plant Disease Diagnosis · Cureus Journal of Computer Science · 2026

Related topics in Agricultural and Biological Sciences