Integrating deep learning applications into agricultural IoT systems faces a serious challenge of balancing the high accuracy of Vision Transformers (ViTs) with the efficiency demands of resource-constrained edge devices. Large transformer models like the Swin Transformers excel in plant disease classification by capturing global-local dependencies. However, their computational complexity (34.1 GFLOPs) limits applications and renders them impractical for real-time on-device inference. Lightweight models such as MobileNetV3 and TinyML would be suitable for on-device inference but lack the required spatial reasoning for fine-grained disease detection. To bridge this gap, we propose a hybrid knowledge distillation framework that synergistically transfers logit and attention knowledge from a Swin Transformer teacher to a MobileNetV3 student model. Our method includes the introduction of adaptive attention alignment to resolve cross-architecture mismatch (resolution, channels) and a dual-loss function optimizing both class probabilities and spatial focus. On the lantVillage-Tomato dataset (18,160 images), the distilled MobileNetV3 attains 92.4% accuracy relative to 95.9% for Swin-L but at an 95% reduction on PC and < 82% in inference latency on IoT devices. (23ms on PC CPU and 86ms/image on smartphone CPUs). Key innovations include IoT-centric validation metrics (13 MB memory, 0.22 GFLOPs) and dynamic resolution-matching attention maps. Comparative experiments show significant improvements over standalone CNNs and prior distillation methods, with a 3.5% accuracy gain over MobileNetV3 baselines. Significantly, this work advances real-time, energy-efficient crop monitoring in precision agriculture and demonstrates how we can attain ViT-level diagnostic precision on edge devices. Code and models will be made available for replication after acceptance.
Agricultural field operations are generally tracked as work records (WR), incorporating data points such as; work type, machine type, timestamped trajectories and field information. WR data which is automatically recorded by modern machinery equipped with Information and Communication Technologies (ICT) can enable efficient farm management decision making. Globally, farmers often rely on aged or legacy farming machinery and manual data recording, which introduces significant labor costs and increases the risk of inaccurate data input. To address this challenge, a field study in Central Japan was conducted to showcase automated data collection by retrofitting legacy farming machinery with low-cost Internet of Things (IoT) devices. For single-purpose vehicles (SPV), which only carry out single work types such as planting, LTE (Long Term Evolution) and Global Navigation Satellite System (GNSS) units were installed to record trajectory data. For multi-purpose vehicles (MPV), such as tractors which perform multiple work types, the configuration settings of these vehicles had to include implements and attachments data. To obtain this data, industry standard LTE-GNSS Bluetooth gateways were fitted onto MPV and low-cost BLE (Bluetooth Low Energy) beacons were attached to implements. After installation, over a seven-month field preparation and planting period 1,623 WR, including 421 WR for SPV and 1,120 WR for MVP, were automatically obtained. For MPV, the WR included detailed configuration settings enabling detection of the specific work types. These findings demonstrate the potential of low cost IoT GNSS devices for precision agriculture strategies to support management decisions in farming operations.
Francisco Zambrano, Abel Herrera, Mauricio Olguín
et al.
The common practice for irrigation management is to apply the water lost by evapotranspiration. However, we could manage the irrigation by monitoring the plant's water status by measuring the stem water potential (Ψs), which is currently costly and time-consuming. The primary goal of this work is to predict the daily spatial variation of Ψs using machine learning models. We measured Ψs in two orchards planted with sweet cherry tree variety Regina, and we monitored 30 trees weekly and biweekly in the central part of Chile, during two seasons, 2022–2023 and 2023–2024, and between October and April. To predict the Ψs, we used the random forest (RF), extreme gradient boosting (XGBoost), and support vector machine (SVM) models. We selected vapor pressure deficit (VPD), reference evapotranspiration (ETo), relative humidity, and temperature as weather predictors. Also, we used as predictors spectral vegetation indices (VIs) and biophysical parameters derived from Sentinel-2. We compared two schemes, one for estimation and another for prediction. We discovered that XGboost and RF worked best for both. The estimation had an R2 of 0.76 and an RMSE of 0.24 MPa. The prediction, on the other hand, had an R2 of 0.59 and an RMSE of 0.36 MPa. The analysis of importance variables reveals that weather predictors, such as VPD, ETo, and temperature, have a higher weight in the model. These are followed by VIs that use short-wave infrared regions, which highlight the moisture stress index (MSI) and the disease and water stress index (DWSI). Future research should tackle the challenge of generalizing the approach, despite its methodological strengths, given that it was based on only two orchards.
Accurate and efficient detection of strawberry runners is crucial for research applications and for developing runner removal solutions in commercial fruit production. This study presented a deep learning-based approach to automate runner detection under field conditions using images collected from diverse platforms. A comprehensive training dataset was assembled, including multiple strawberry varieties, growth stages, and seasons. Images were captured using three approaches: ground imaging (GI) at 0.5 m above ground level (AGL) in a forward view, and aerial imaging at 5 m AGL (AI5) and 10 m AGL (AI10), both in a nadir view. These platforms provided strawberry plant images of varying resolutions to train YOLO-based deep convolutional neural network (DCNN) models for runner detection and segmentation. Models were trained on platform-specific datasets (GI, AI5, and AI10) as well as on an integrated dataset (GI+AI5+AI10). Validation results revealed that detection and segmentation models trained on the integrated dataset outperformed those trained on platform-specific individual datasets, demonstrating stronger generalization across diverse imaging conditions. The detection model achieved F1-scores of 0.79, 0.81, and 0.82 on the GI, AI5, and AI10 validation datasets, respectively, outperforming the corresponding segmentation models and highlighting the benefit of multi-source training. Notably, for runner detection, aerial imaging at 5 m AGL achieved a strong balance between detection accuracy and imaging efficiency, with an F1-score of 0.81, outperforming both ground-based and higher-altitude aerial imagery. This study improved the automation of strawberry runner detection, provided a publicly accessible dataset for training DCNN models for strawberry plant parts detection, and highlighted mid-altitude aerial imaging as a practical solution for high-throughput runner detection.
Pesticide precision spraying and efficient deposition is an important development direction of smart agriculture. Aiming at the problems of low pesticide spraying efficiency and severe pesticide loss in unstructured orchards in hilly and mountainous areas, this study proposes an automatic target spray control method. A tracked orchard sprayer based on millimeter-wave radar is designed to address these issues. The information transmission between millimeter wave radar, controller and sprayer are realized, and automatic target spray operation of ''Walking-Sensing-Spraying'' are realized. Based on the improved self-adaptive DBSCAN clustering algorithm, the improved self-adaptive Alpha_Shape algorithm (a surface reconstruction algorithm) and the least squares circle fitting, the three-dimensional reconstruction and parameter extraction of the target canopy were realized. The results showed that the average relative errors of plant height, canopy width and volume after correction were 1.51 %, 1.96 % and 3.24 %, respectively. The maximum absolute error is 9.59 cm, 5.96 cm and 0.22 m3. The millimeter-wave radar point cloud can effectively characterize the plant height, canopy width and volume information of the target canopy, and meet the detection accuracy requirements of target spraying. Field experiment results show that the spray coverage under t automatic target spray meets the needs of orchard pest control, the application of pesticides is reduced by 36.12 %, which achieves the purpose of increasing efficiency, reducing application and precise application. Meanwhile, it can also provide methodological reference for other research on automatic target operation and other fields of automatic target spray technology.
Hanaa A. El-Hamshary, Lamiaa I. El-Nawasany, Osama Magouz
et al.
The present study aimed to develop a functional yogurt enriched with microencapsulated hydroethanolic extract of Vitex agnus-castus L. fruit. The extract showed potent antioxidant activity (85.98 % DPPH scavenging at 0.5 mg/mL; FRAP absorbance of 0.51 at 700 nm) and selective cytotoxicity against human prostate cancer (PC3), human epidermoid carcinoma (A431), human Caucasian breast adenocarcinoma (MCF7), and human hepatocellular carcinoma (HepG2) cells (IC₅₀: 38.27–80.97 µg/mL), with minimal impact on normal BJ1 fibroblasts. High-performance liquid chromatography with diode array detection (HPLC-DAD) analysis revealed catechin, quercetin, gallic acid, and chlorogenic acid as major phenolics. In silico molecular docking supported the observed bioactivity by illustrating interactions of key phytochemicals with cancer-related targets. To enhance stability and bioavailability, V. agnus-castus extract was encapsulated via complex coacervation using gum Arabic and whey protein. The resulting microcapsules demonstrated favorable encapsulation efficiency, physicochemical characteristics, and resistance to simulated gastrointestinal conditions. Incorporation into yogurt enhanced its nutritional value, texture, and sensory attributes. Overall, V. agnus-castus extract-loaded microcapsules show promise as functional food additives with antioxidant and anticancer potential.
Yu-Pin Lin, Pei-Chen Lin, Shafira Wuryandani
et al.
Managing food–energy–water (FEW) resources in the face of climate change presents significant challenges, particularly owing to a limited understanding of the spatiotemporal dynamics of ecosystem service (ES)-driven FEW sustainability. This study presents a novel approach to evaluate FEW sustainability under climate and land use changes by utilizing ES values instead of the supply-demand of FEW resources for the Zhuoshui Watershed in Taiwan. We integrate the Conversion of Land Use and its Effects on the Small Regional Extensions model and three general circulation models under the RCP 8.5 scenario to simulate land use change under climate change. Associated ES changes are estimated using the InVEST model, identifying ES hotspots with Local Indicators of Spatial Association and quantifying FEW sustainability through geographically weighted regression methods (R2 = 0.65). Land use changes significantly influenced ES provision, with conservation areas enhancing ecological resilience and water sustainability. ES hotspots were associated with higher water security but faced trade-offs in food and energy subsystems. Changes in land use and associated ES strongly influenced FEW sustainability, particularly in conservation areas, by promoting ecological resilience and sustainable resource management. Key measures include ecosystem-based spatial planning and strategies for FEW sustainability under climate change. These measures help maintain enhanced ES provision, improve sustainability projections, and support evidence-based policies for long-term environmental sustainability. Hotspot regions exhibit higher water sustainability but will be increasingly pressured regarding food and energy resources.
Tiffany W. O’Connor, Damian Collins, Andrew J. Read
et al.
Quantifying the infectious titre of preparations containing rabbit haemorrhagic disease virus (RHDV) is an essential virological technique during RHDV research. The infectious titre of an RHDV preparation is determined using a bioassay to identify the endpoint dilution at which 50% of rabbits become infected (RID<sub>50</sub>). Previous publications have briefly described the method for estimating the infectious titre of RHDV preparations by challenging rabbits with 10-fold serial dilutions. However, these descriptions lack the critical considerations for a standardised method to estimate RID<sub>50</sub>. These details are presented here, along with a comparison between the Reed–Muench, Dragstedt–Behrens, Spearman–Kärber, and probit regression methods for calculating the RID<sub>50</sub>. All the statistical approaches demonstrated a high level of agreement in calculating the RID<sub>50</sub>. To help assess the precision of the estimated infectious titre, the improved Spearman–Kärber and probit regression methods provide the 95% confidence intervals. The method outlined improves the accuracy of results when undertaking studies of pathogenicity, host resistance, and the production of vaccines against RHDV.
Amany Behairy, Walaa El-Houseiny, Abdallah Tageldein Mansour
et al.
IntroductionTitanium dioxide nanoparticles (TDNPs) are widely used in food industries, agricultural and consumer products, and diagnostic purposes, leading to their potential release into aquatic environments and associated physiological risks to non-target aquatic organisms, particularly fish. α-Sitosterol (STL), a phytosterol, acts by enhancing antioxidant defenses and modulating inflammatory signaling pathways. Hence, this study investigated whether dietary STL can protect Nile tilapia (Oreochromis niloticus) from TDNPs-induced toxicity during a 60-day exposure.MethodsIn this study, 300 Nile tilapia were allocated into four groups. The control group received a basal diet, the STL group was fed a diet supplemented with 80 mg STL/kg, the TDNPs group was exposed to 10 mg/L of TDNPs in water, and the TDNPs + STL group was exposed to TDNPs and fed the STL-supplemented diet.ResultsDietary STL supplementation markedly improved growth performance, with increases of 33%–60% in final body weight, weight gain, and daily growth rate, and a 29% reduction in feed conversion ratio compared to TDNPs-exposed fish. STL supplementation also restored hematological parameters altered by TDNPs exposure, including significant recovery of red blood cells, hemoglobin, packed cell volume, and white blood cells, thereby reversing macrocytic normochromic anemia and leukopenia. Furthermore, STL significantly decreased elevated serum alanine aminotransferase, aspartate aminotransferase, alkaline phosphatase, urea, and creatinine levels induced by TDNPs, and normalized lipid profiles by reducing total cholesterol, triglycerides, and low-density lipoprotein cholesterol, while elevating high-density lipoprotein cholesterol. STL-fed fish also exhibited significant reductions in stress biomarkers (glucose and cortisol) and enhanced innate immune responses, as evidenced by higher lysozyme, complement 3, nitric oxide, nitro blue tetrazolium, and phagocytic activity. Antioxidant status was strengthened through increased superoxide dismutase, catalase, and glutathione peroxidase activities and reduced malondialdehyde levels. At the molecular level, STL supplementation downregulated endoplasmic reticulum stress-related genes (chop, jnk, xbp-1, and perk), while upregulating autophagy-related genes (beclin-1 and lc3-ii) and downregulating mtor and p62. Histological analysis confirmed STL’s protective effects, showing marked recovery of intestinal, hepatic, renal, and splenic structures.ConclusionThese findings demonstrated that STL confers multi-level protection against TDNPs-induced oxidative, metabolic, and cellular stress, highlighting its potential as a functional dietary supplement for mitigating nanotoxicity in aquaculture.
Science, General. Including nature conservation, geographical distribution
Maarten Perneel, Ines Adriaens, Ben Aernouts
et al.
Over the past decade, studying animal behaviour with the help of computer vision has become more popular. Replacing human observers by computer vision lowers the cost of data collection and therefore allows to collect more extensive datasets. However, the majority of available computer vision algorithms to study animal behaviour is highly tailored towards a single research objective, limiting possibilities for data reuse. In this perspective, pose-estimation in combination with animal tracking offers opportunities to yield a higher level representation capturing both the spatial and temporal component of animal behaviour. Such a higher level representation allows to answer a wide variety of research questions simultaneously, without the need to develop repeatedly tailored computer vision algorithms. In this paper, we first cope with several weaknesses of current pose-estimation algorithms and thereafter introduce KeySORT (Keypoint Simple and Online Realtime Tracking). KeySORT deploys an adaptive Kalman filter to construct tracklets in a bounding-box free manner, significantly improving the temporal consistency of detected keypoints. In this paper, we focus on pose estimation in cattle, but our methodology can easily be generalised to any other animal species. Our algorithm is able to detect up to 80% of the ground truth keypoints with high accuracy, with only a limited drop in performance when daylight recordings are compared to nightvision recordings. Moreover, by using KeySORT to construct skeletons, the temporal consistency of generated keypoint coordinates was largely improved, offering opportunities with regard to automated behaviour monitoring of animals.
As global climate change intensifies with unprecedented urgency, nations worldwide have increasingly adopted market-based environmental regulatory instruments to advance carbon reduction objectives. In 2017, China launched energy rights trading pilots, thereby providing a crucial policy instrument for controlling total energy consumption at its source. However, the specific impacts and transmission pathways through which this system influences corporate carbon reduction behavior remain insufficiently explored through rigorous empirical investigation. Drawing upon panel data from heavy-polluting companies listed on the Shanghai and Shenzhen A-share markets, this study employs a difference-in-differences methodology to identify the causal effects of energy rights trading systems on corporate carbon reduction. Our findings reveal that energy rights trading systems significantly reduce corporate carbon emission intensity, generating pronounced emission reduction effects. Further mechanism analysis demonstrates that this system operates through two principal pathways: first, by promoting increased green investment among enterprises, whereby short-term emission reductions are achieved through procurement of energy-saving equipment and environmental protection facilities, and second, by stimulating corporate green technological innovation, whereby long-term sustainable emission reductions are realized through the development of energy-saving technologies and clean processes. Additionally, the research reveals that enterprises with lower financing constraints and stronger supply chain bargaining power respond more actively to policy implementation, with policy effects exhibiting significant heterogeneity. This study not only enriches the theoretical understanding of market-based environmental regulatory policy effects but also provides crucial empirical evidence for improving the energy rights trading system design and enhancing policy implementation effectiveness, thereby offering important policy insights for promoting corporate green transformation and achieving “dual carbon” objectives.
The realization of FDI and DDI from January to December 2022 reached Rp1,207.2 trillion. The largest FDI investment realization by sector was led by the Basic Metal, Metal Goods, Non-Machinery, and Equipment Industry sector, followed by the Mining sector and the Electricity, Gas, and Water sector. The uneven amount of FDI investment realization in each industry and the impact of the COVID-19 pandemic in Indonesia are the main issues addressed in this study. This study aims to identify the factors that influence the entry of FDI into industries in Indonesia and measure the extent of these factors' influence on the entry of FDI. In this study, classical assumption tests and hypothesis tests are conducted to investigate whether the research model is robust enough to provide strategic options nationally. Moreover, this study uses the ordinary least squares (OLS) method. The results show that the electricity factor does not influence FDI inflows in the three industries. The Human Development Index (HDI) factor has a significant negative effect on FDI in the Mining Industry and a significant positive effect on FDI in the Basic Metal, Metal Goods, Non-Machinery, and Equipment Industries. However, HDI does not influence FDI in the Electricity, Gas, and Water Industries in Indonesia.
Dhanushka Balasingham, Sadeesha Samarathunga, Gayantha Godakanda Arachchige
et al.
The advancements in precision agriculture are vital to support the increasing demand for global food supply. Precision spot spraying is a major step towards reducing chemical usage for pest and weed control in agriculture. A novel spot spraying algorithm that autonomously detects weeds and performs trajectory planning for the sprayer nozzle has been proposed. Furthermore, this research introduces a vision-based autonomous navigation system that operates through the detected crop row, effectively synchronizing with an autonomous spraying algorithm. This proposed system is characterized by its cost effectiveness that enable the autonomous spraying of herbicides onto detected weeds.
Sudi Murindanyi, Joyce Nakatumba-Nabende, Rahman Sanya
et al.
The increasing popularity of Artificial Intelligence in recent years has led to a surge in interest in image classification, especially in the agricultural sector. With the help of Computer Vision, Machine Learning, and Deep Learning, the sector has undergone a significant transformation, leading to the development of new techniques for crop classification in the field. Despite the extensive research on various image classification techniques, most have limitations such as low accuracy, limited use of data, and a lack of reporting model size and prediction. The most significant limitation of all is the need for model explainability. This research evaluates four different approaches for crop classification, namely traditional ML with handcrafted feature extraction methods like SIFT, ORB, and Color Histogram; Custom Designed CNN and established DL architecture like AlexNet; transfer learning on five models pre-trained using ImageNet such as EfficientNetV2, ResNet152V2, Xception, Inception-ResNetV2, MobileNetV3; and cutting-edge foundation models like YOLOv8 and DINOv2, a self-supervised Vision Transformer Model. All models performed well, but Xception outperformed all of them in terms of generalization, achieving 98% accuracy on the test data, with a model size of 80.03 MB and a prediction time of 0.0633 seconds. A key aspect of this research was the application of Explainable AI to provide the explainability of all the models. This journal presents the explainability of Xception model with LIME, SHAP, and GradCAM, ensuring transparency and trustworthiness in the models' predictions. This study highlights the importance of selecting the right model according to task-specific needs. It also underscores the important role of explainability in deploying AI in agriculture, providing insightful information to help enhance AI-driven crop management strategies.
Karthik Muthineni, Alexander Artemenko, Josep Vidal
et al.
The fifth generation (5G) mobile communication technology integrates communication, positioning, and mapping functionalities as an in-built feature. This has drawn significant attention from industries owing to the capability of replacing the traditional wireless technologies used in industries with 5G infrastructure that can be used for both connectivity and positioning. To this end, we identify the Automated Guided Vehicle (AGV) as a primary use case to benefit from the 5G functionalities. Given that there have been various works focusing on 5G positioning, it is necessary to analyze the existing works about their applicability with AGVs in industrial environments and provide insights to future research. In this paper, we present state of the art in 5G-based positioning, with a focus on key features, such as Millimeter Wave (mmWave) system, Massive Multiple Input Multiple Output (MIMO), Ultra-Dense Network (UDN), Device-to-Device (D2D) communication, and Reconfigurable Intelligent Surface (RIS). Moreover, we present the shortcomings in the current state of the art. Additionally, we propose enhanced techniques that can complement the accuracy of 5G-based positioning in controlled industrial environments.
Irrigation decision-making has evolved from solely depending on farmers’ decisions taken based on the visual analysis of field conditions to making decisions based on crop water need predictions generated using machine learning (ML) techniques. This paper reviews ML related articles to discuss how ML has been used to enhance irrigation decision making. We reviewed 16 studies that used ML approaches for irrigation scheduling prediction and decision-making focusing on the input features, algorithms used and their applicability in real world conditions. ML performances in terms of accuracy, water conservation compared to fixed or threshold-based methods are discussed along with modeling performances. Informed by the 16 research studies, we assessed constraints to the adoption of ML in irrigation decision making at field scale, which include limited data availability coupled with data sharing constraints, and a lack of uncertainty quantification as well as the need for physics informed ML based irrigation scheduling models. To address these limitations, we discussed approaches in future research such as integrating process-based models with ML, incorporating expert knowledge into the modeling procedure, and making data and tools Findable, Accessible, Interoperable, and Reusable (FAIR). These approaches will improve ML modeling outcomes and boost the availability of farm-related data and tools for FAIRer data-driven applications of irrigation modeling.
Jéssica Elaine Silva, Wender Henrique Batista da Silva, Marcelo Araújo Junqueira Ferraz
et al.
The use of drones in coffee farming has emerged as a solution to optimize phytosanitary and nutritional management, particularly in crops with dense canopies. Effective drone spraying requires precise adjustments to parameters like spray volume and flight speed to maintain application efficiency as the coffee plants grow. This study investigated the efficiency of drone spraying across varying spray volumes and flight speeds in coffee plants of two distinct ages, focusing on droplet deposition within the upper, middle, and lower thirds of the plant canopy. The experiment was conducted in coffee plantations aged 2.5 and 6.5 years, with treatments consisting of four spray volumes (8, 12, 16, and 20 L ha⁻¹) and two flight speeds (12 and 15 km h⁻¹), each replicated three times. Water-sensitive papers were placed in the three canopy layers to assess droplet deposition. In younger coffee plants, the position of the water-sensitive paper showed no significant differences in droplet distribution, while in older plants, larger droplets were predominantly found in the upper third of the canopy. Spray volume and flight speed influenced the droplet spectrum, with volumes of 8 and 16 L ha⁻¹ at 12 km h⁻¹ producing larger droplets in 2.5-year-old plants, whereas 20 L ha⁻¹ at 12 km h⁻¹ resulted in larger droplets in 6.5-year-old plants. These findings underscore the importance of calibrating drone parameters based on plant age, product type, and target location, as spray volume and flight speed significantly affect product distribution and canopy penetration.
Using deep learning techniques, we introduce a novel measure for production process heterogeneity across industries. For each pair of industries during 1990-2021, we estimate the functional distance between two industries' production processes via deep neural network. Our estimates uncover the underlying factors and weights reflected in the multi-stage production decision tree in each industry. We find that the greater the functional distance between two industries' production processes, the lower are the number of M&As, deal completion rates, announcement returns, and post-M&A survival likelihood. Our results highlight the importance of structural heterogeneity in production technology to firms' business integration decisions.
Industry 4.0 operates based on IoT devices, sensors, and actuators, transforming the use of computing resources and software solutions in diverse sectors. Various Industry 4.0 latency-sensitive applications function based on machine learning to process sensor data for automation and other industrial activities. Sending sensor data to cloud systems is time consuming and detrimental to the latency constraints of the applications, thus, fog computing is often deployed. Executing these applications across heterogeneous fog systems demonstrates stochastic execution time behavior that affects the task completion time. We investigate and model various Industry 4.0 ML-based applications' stochastic executions and analyze them. Industries like oil and gas are prone to disasters requiring coordination of various latency-sensitive activities. Hence, fog computing resources can get oversubscribed due to the surge in the computing demands during a disaster. We propose federating nearby fog computing systems and forming a fog federation to make remote Industry 4.0 sites resilient against the surge in computing demands. We propose a statistical resource allocation method across fog federation for latency-sensitive tasks. Many of the modern Industry 4.0 applications operate based on a workflow of micro-services that are used alone within an industrial site. As such, industry 4.0 solutions need to be aware of applications' architecture, particularly monolithic vs. micro-service. Therefore, we propose a probability-based resource allocation method that can partition micro-service workflows across fog federation to meet their latency constraints. Another concern in Industry 4.0 is the data privacy of the federated fog. As such, we propose a solution based on federated learning to train industrial ML applications across federated fog systems without compromising the data confidentiality.