The researchers evaluated two commercially cultivated cucumber varieties, CU 999 and Monalisa, at three plant spacings—75 × 25, 75 × 50 and 75 × 75 cm—and four NPK rates of 0, 100, 200 and 300 kg ha−1. The treatments were arranged in a split-split-plot randomized complete block design with three replications per trial. Each sub-subplot covered 6 m², with 0.5 m alleys between plots and 1 m between blocks. Fruit was harvested at commercial maturity, defined as 10–12 days after anthesis, with uniform dark-green skin and a length of 14–18 cm. Five randomly selected fruits from the middle two rows of each sub-subplot, excluding border plants, were taken for laboratory analysis.

CU 999 generally produced firmer fruit, while Monalisa consistently had greener skin. In the second 2019 trial, firmness averaged 1.35 kg for CU 999 compared with 1.15 kg for Monalisa; in the second 2020 trial, the corresponding values were 1.20 and 1.05 kg. Wider spacing also supported firmness, with fruits at 75 × 75 cm reaching 1.35 kg in the second 2019 trial compared with 1.07 kg at 75 × 25 cm. Fertilizer rate also affected firmness, but the response was not consistently linear: the highest mean occurred at 100 kg ha−1 in the first 2019 trial, at the unfertilized control in the second 2019 trial, and at 300 and 200 kg ha−1 in the two 2020 trials. In the first 2020 trial, the firmest treatment combination was CU 999 at 75 × 75 cm with 300 kg ha−1, at 1.46 kg.

Monalisa’s advantage in green coloration was more consistent than any fertilizer effect. Its CIELAB a* values were more negative than those of CU 999 in all four trials, indicating greener skin. In the second 2019 trial, the values were −12.13 for Monalisa and −10.54 for CU 999; in the first 2020 trial, they were −7.94 and −3.52, respectively. Spacing altered skin color significantly in every trial, although wider spacing produced greener fruit in only two trials. A rate of 200 kg ha−1 generally corresponded to greener skin than lower or higher rates, but this pattern was not uniform. Significant three-way interactions among variety, spacing and fertilizer were recorded for both firmness and color in all four trials, showing that the best combination depended on season and the trait under examination.

The clearest nutritional response was associated with the highest fertilizer rate. NPK at 300 kg ha−1 significantly increased vitamin A in every trial, reaching 35.93 µg/100 g in the second 2019 trial compared with 30.39 µg/100 g in the unfertilized control. It also produced the highest vitamin K values in all four trials, including 19.05 µg/100 g versus 15.84 µg/100 g under the control in the first 2019 trial. Vitamin C responded significantly to fertilizer rate in each trial, but not in a uniformly dose-dependent fashion: the highest mean occurred at 100 kg ha−1 in the first 2019 trial, 300 kg ha−1 in the second, and 200 and 300 kg ha−1 in the two 2020 trials. Monalisa generally had slightly more vitamin A than CU 999, while the varieties did not differ significantly in vitamin K in most trials.

Plant spacing was especially influential for vitamin K and vitamin C. The 75 × 75 cm treatment produced the highest vitamin K concentration in every trial, including 17.55 µg/100 g compared with 16.50 µg/100 g at 75 × 25 cm in the second 2019 trial. It also yielded the highest mean vitamin C concentration in the second 2019 trial and both 2020 trials, while the first 2019 trial instead favored 75 × 50 cm. These observations are consistent with the broader pattern in which reduced competition at wider spacing was associated with more concentrated fruit constituents. The relationship was not simply a matter of maximizing every vitamin at once: riboflavin, identified in the study as vitamin B2, remained uniformly low at 0.03–0.04 mg/100 g, and the 300 kg ha−1 fertilizer rate increased it significantly only in the two 2020 trials.

Fruit composition showed the same tension between output and concentration. The closest spacing maximized per-hectare yield, but its fruit contained less vitamin C in the study’s interpretation, consistent with nutrient dilution at high plant density. Wider spacing generally increased dry matter. NPK at 300 kg ha−1 gave the highest dry matter in three of four trials, including 5.27 g/100 g in the first 2020 trial, compared with 3.46 g/100 g in the control. The 75 × 75 cm treatment generally produced higher dry matter, while the 75 × 50 cm treatment sometimes gave the highest carbohydrate level. Fiber ranged from 0.25 to 0.50 g/100 g, with wider spacing and higher NPK rates generally associated with higher values. Fat remained low throughout, at 0.10–0.15 g/100 g, and did not differ significantly between varieties.

The multivariate analysis summarized the combined pattern. CU 999 at 75 × 75 cm with 300 kg ha−1 achieved the highest TOPSIS closeness coefficient, 0.768, indicating the strongest overall position across the measured postharvest and nutritional criteria. By contrast, CU 999 at 75 × 25 cm with the same fertilizer rate was the yield-maximizing recommendation. Principal component analysis placed vitamin content, dry matter and firmness together along the first component, while moisture and proximate fractions dominated the second. Treatments receiving 300 kg ha−1 consistently shifted toward higher scores on the first component. The moisture content was strongly inversely correlated with dry matter, with r = −0.93, and negatively associated with most vitamins. Firmness correlated positively with dry matter at r = 0.72 and with vitamin K at r = 0.68.

These results show why a single “best” treatment cannot summarize the experiment. The study analyzed each season independently because planting date, rainfall and soil conditions differed, and the four trials did not support a combined multi-environment analysis without explicit site-year covariates. ANOVA and Fisher’s least significant difference were used for treatment comparisons, while correlation, principal component analysis and TOPSIS were used to examine the combined profile. The 75 × 75 cm and 300 kg ha−1 combination should therefore be interpreted as a quality-oriented option rather than a definitive commercial recommendation. It uses fewer plants per unit area and requires a high fertilizer input, so profitability, fertilizer efficiency and environmental risk remain unresolved.

The findings nevertheless offer a clear experimental distinction for southwestern Nigerian cucumber production: close spacing favored volume, whereas wider spacing repeatedly favored dry matter and selected vitamins; CU 999 favored firmness, and Monalisa favored green skin color. The authors also identify limitations that constrain interpretation. Trials were conducted at one site over two years, economic feasibility was not tested, and vitamins A and K were quantified spectrophotometrically rather than by the more selective HPLC method. The study therefore demonstrates a management-linked trade-off in cucumber quality under tropical open-field conditions, while further location-based, economic and environmental studies are needed before the highest NPK rate can be presented as a broadly optimal production choice.

Subject of Research: Field study of cucumber variety, spacing and NPK effects on yield, postharvest quality and nutritional composition

Article Title: Variety, plant spacing, and NPK fertilizer rate influence postharvest quality and nutritional composition of cucumber fruits in Southwest Nigeria

Article References: Falaiye, O. O., Kasali, A. O., Adeyemi, N. O., Oni, O. E., Somoye, R. A., Falaiye, O. M., & Adekunle, E. A. (2026). Variety, plant spacing, and NPK fertilizer rate influence postharvest quality and nutritional composition of cucumber fruits in Southwest Nigeria. Discover Plants, 3(1), Article 473. https://doi.org/10.1007/s44372-026-00957-7

Image Credits: AI Generated

DOI: 10.1007/s44372-026-00957-7

Keywords: cucumber, crop management, plant spacing, NPK fertilizer, food nutrition, postharvest quality, Nigeria, yield trade-offs, CU 999, Monalisa, Variety, plant

The logic behind distributed machine learning is deceptively simple. In the traditional model, all data from sensors, drones, and farm management systems is shipped to a central server where a single model is trained. That architecture carries well-known weaknesses: a single point of failure, bandwidth latency, and acute privacy risks, since consolidating records about crop production, animal health, and economic yields creates a tempting target for massive breaches. Distributed learning inverts this model. Each participant trains a model locally on their own devices, and only the learned parameters, not the raw data, travel across the network. A central server, or in some architectures no server at all, then combines these updates into a shared global model. The result is enhanced privacy, security, efficiency, and scalability, qualities that matter enormously in a sector where farmers increasingly worry about data asymmetry and how agricultural technology providers store and exploit their information.

To capture this rapidly evolving landscape, the researchers followed the systematic literature review guidelines of Kitchenham and colleagues, ensuring the process was replicable, unbiased, thorough, and transparent. They searched six major databases, including IEEE Xplore, ACM Digital Library, Scopus, ScienceDirect, SpringerLink, and Wiley Online Library, covering publications from 2015 to 2025. The initial automated search returned 1,864 articles. After screening against strict exclusion criteria, only journal articles were considered, and titles, abstracts, and full texts were evaluated, 43 studies survived. A quality assessment on an eight-point scale, scored independently by all three authors and reconciled through consensus meetings, trimmed the set to 41 primary studies. Notably, no study published before 2020 met the criteria, a striking signal that distributed learning in agriculture is a genuinely recent phenomenon that has accelerated sharply in the past five years.

The review’s findings reveal a field with clear centers of gravity. Crop management dominates the application landscape, appearing far more frequently than any other domain, with livestock monitoring, soil and water management, and general smart farming frameworks trailing behind. Within these domains, the researchers identified seventeen distinct tasks, ranging from disease prediction and classification, crop classification, and crop recommendation to weed segmentation, yield prediction, resource optimization, evapotranspiration estimation, bird detection, milk quality prediction, and even prawn growth inspection. Disease prediction and classification emerged as the most common objective, reflecting the enormous value of catching plant pathogens early. One study, a federated explainable AI framework, tackled the widest range of tasks simultaneously, spanning disease prediction, crop health monitoring, yield prediction, resource optimization, and anomaly detection.

Among the distributed learning paradigms, federated learning stands overwhelmingly supreme, appearing across every one of the seventeen identified tasks. It is followed by decentralized peer-to-peer systems, hierarchical distributed learning, and blockchain-based approaches, with two studies proposing hybrids that combine federated learning with blockchain for applications such as pest detection. Hierarchical models insert intermediate edge servers between local devices and the cloud to reduce communication overhead and latency. Decentralized peer-to-peer architectures remove the central coordinator entirely, with nodes exchanging updates only with their direct neighbors, making performance highly dependent on network topology. Blockchain-based systems replace the central server with an immutable distributed ledger, using smart contracts to automatically verify, validate, and aggregate model updates. Each approach embodies a trade-off between control and resilience: centralized systems are simpler and faster to manage but collapse if the main server fails, while decentralized designs keep running even when individual devices drop offline, at the cost of greater network traffic and slower synchronization.

The architectural details expose an interesting tension. Although federated learning distributes the data, most implementations still rely on centralized client-server topologies, which appeared in 27 of the reviewed studies, compared with 12 using hierarchical tree architectures and only 2 using peer-to-peer or mesh designs. Governance follows architecture: centralized authority and server-dictated models dominate, in which a single controlling entity manages data flows and decision-making. Yet the review uncovered notable exceptions, including a client-server system that adopted collaborative, consensus-based governance by combining its architecture with blockchain, requiring agreement among local nodes before updating the global model. Aggregation strategies cluster around Federated Averaging, used in 27 studies, followed by custom weighted averaging and Federated Proximal Aggregation, with more exotic variants such as SCAFFOLD, FedBN, and QFedAvg appearing once or twice each.

On the data side, images reign as the most common input, which aligns with the dominance of deep learning algorithms, particularly convolutional neural networks such as YOLO, ResNet, VGG, MobileNet, and Inception, alongside transformers like Vision Transformer and Swin Transformer. Classical machine learning methods, including support vector machines, random forests, decision trees, and k-nearest neighbors, handle tabular, sensor, and time-series data. Implementation leans on a familiar toolkit: TensorFlow leads with 12 studies, followed by Keras, Flower, and PyTorch, with TensorFlow Federated, PySyft, MATLAB, scikit-learn, iFogSim, Horovod, and Open MPI filling out the ecosystem. Hardware ranges from Raspberry Pi, Jetson, and Arduino edge devices to NVIDIA GPUs and cloud platforms, underscoring that image-heavy distributed learning still demands serious computing resources.

Perhaps the most valuable contribution of the review is its honest accounting of what still goes wrong. Three challenges recur across the literature. Privacy leakage persists because shared model updates, gradients and weights, carry mathematical footprints of private data, enabling sophisticated attacks such as gradient reconstruction, model inversion, and membership inference. Communication bottlenecks arise because continuous model updates from thousands of devices can strain network transmission far beyond local computing speeds, stretching training from hours to weeks, especially under unstable bandwidth and intermittent rural connectivity. Data heterogeneity, the notorious non-IID problem, means each farm’s data is statistically unique and skewed, so a global model can underperform local models and quietly favor majority clients, creating unfair bias against smaller participants.

The solutions documented in the literature are correspondingly diverse. Against privacy threats, researchers deploy differential privacy, which injects noise into data and updates, along with homomorphic encryption, secure multi-party computation, gradient encryption, and blockchain frameworks augmented with decentralized identifiers and verifiable credentials. To tame communication costs, the field turns to gradient compression, quantization, binary compression, knowledge distillation, adaptive and asynchronous transmission, lightweight architectures, and model pruning. For non-IID data, promising remedies include advanced aggregation algorithms such as FedProx and SCAFFOLD, adaptive weighting, data refinement techniques like resampling, zero-padding, and augmentation, and ensemble or prototype-based learning to boost robustness. The authors conclude by proposing a multi-dimensional taxonomy, categorizing systems by domain, architecture, learning method, infrastructure, and challenges, and call for reference architectures and for extending distributed learning beyond crops into livestock monitoring and precision aquaculture. As farming becomes ever more data-driven, this review makes clear that the future of agricultural AI may depend less on how much data is collected, and more on how intelligently, and safely, it is shared.

Subject of Research: Systematic literature review of distributed machine learning methods, applications, architectures, and challenges in precision agriculture

Article Title: Distributed machine learning in precision agriculture: methods, applications, and challenges

Article References: Krisnawijaya, N. N. K., Tekinerdogan, B., & Catal, C. (2026). Distributed machine learning in precision agriculture: methods, applications, and challenges. Smart Agricultural Technology, 15, Article 102527. https://doi.org/10.1016/j.atech.2026.102527

Image Credits: AI Generated

DOI: 10.1016/j.atech.2026.102527

Keywords: distributed machine learning, federated learning, precision agriculture, smart farming, data privacy, crop management, non-IID data, edge computing, blockchain, systematic literature review, differential privacy, agricultural IoT