The scale of the evidence base is striking. Following the PRISMA 2020 reporting framework, the team searched ScienceDirect, PubMed, IEEE Xplore, Google Scholar, and MDPI for English-language peer-reviewed work published between January 2021 and December 2025, retrieving 13,926 records. After deduplication, title and abstract screening, and full-text eligibility assessment of 198 articles, 95 studies survived the cut. Quality ran high: 51.58 percent of included papers appeared in first-quartile journals and another 36.84 percent in second-quartile venues, with Journal Citation Reports impact factors ranging from 2.2 to 14.0. The topical map shows where the intelligence is concentrating. Environmental monitoring accounted for 23.15 percent of publications, greenhouse control and protected cultivation for 21.05 percent, disease and pest detection for 21.05 percent, hydroponics for 16.84 percent, and yield prediction with production forecasting for 13.68 percent. Methodologically, classical machine learning anchored 25.26 percent of the studies, IoT and AIoT sensor systems 22.11 percent, and deep learning and computer vision 13.68 percent, a sign that pragmatic, sensor-driven systems currently outweigh frontier models in the field.

At the foundation sits machine learning, in which computers learn patterns from data rather than following explicitly programmed rules. The review organizes these algorithms into supervised, unsupervised, and reinforcement learning families and shows how workhorse models such as decision trees, support vector machines, random forests, and artificial neural networks now digest soil nutrient profiles, fertilizer inputs, and crop characteristics to make site-specific recommendations. In one line of work, random forests, support vector machines, and neural networks predicted optimal water requirements for maize, soybean, and tomato across diverse climates and soils, outperforming conventional irrigation scheduling. Transparency is emerging as a design priority. When researchers applied SHAP and LIME, techniques that reveal which variables drive a model’s predictions, to 15 greenhouse-grown cabbage plants tracked over an 85-day period, they could see that leaf count and plant height strengthened forecasts of nitrogen, phosphorus, and potassium levels while days after planting and average leaf area weakened them. Such explainable AI, the authors argue, is becoming essential for earning the trust of farmers expected to act on algorithmic advice.

Above that layer operates deep learning, multi-layer neural networks that automatically discover hierarchical features in raw data without hand-engineered rules. Convolutional neural networks, the workhorses of agricultural computer vision, learn to recognize edges, textures, and lesions in images streaming from cameras, drones, satellites, and field sensors, enabling real-time monitoring of plant health, growth analysis, and yield estimation. The review documents multimodal systems that fuse hyperspectral and X-ray imaging to predict seed viability non-destructively, alongside architectures such as VGG16, YOLO, Mask R-CNN, and transformer-based models that classify weeds, detect plant stress, and recognize growth stages. The techniques even reach into farm physics: one team built a three-dimensional computational fluid dynamics model to optimize airflow velocity, temperature distribution, and relative humidity across the cultivation trays of an indoor vertical farm, reporting improved airflow uniformity, better thermal control, and higher energy efficiency. Hybrid designs push further still. Convolutional layers coupled with bidirectional long short-term memory networks capture spatial patterns and temporal dynamics together, and a grey-wolf-optimized BiLSTM model forecast greenhouse temperatures with a coefficient of determination of 0.97.

Binding these computational layers together is the Internet of Things, networks of sensors, microcontrollers, and communication modules that measure soil moisture, temperature, humidity, pH, electrical conductivity, light intensity, and nutrient levels, then stream the data over Wi-Fi, LoRa, Zigbee, or cellular links to cloud and edge computing platforms. When AI sits atop those streams, researchers call it AIoT: sensors feed predictive models, models command actuators such as pumps, valves, LED arrays, and climate systems, and the farm adjusts itself in real time. The review catalogs the hardware ecosystem, from Arduino, ESP32, and Raspberry Pi controllers to LoRaWAN gateways and mobile dashboards. One LoRaWAN-based subsurface drip irrigation system, triggered when soil moisture fell to 12 percent, enhanced tomato seedling growth while cutting water use by 10 percent and lifting yield to 1,243 grams per plant. Edge computing is shrinking latency and energy costs further, and digital twins, virtual replicas of physical farms, now let growers simulate crop growth, microclimate behavior, and management scenarios before committing a single resource in the real world.

Nowhere is the transformation more visible than in controlled-environment agriculture, the umbrella term for greenhouses, vertical farms, and plant factories where light, temperature, humidity, and nutrition are engineered rather than endured. The review describes greenhouses that have evolved into quasi-autonomous organisms: one IoT-enabled system used artificial neural networks and regression models combined with image analysis to predict lettuce growth, harvest timing, and crop quality from environmental variables, while another deployed long short-term memory networks to forecast actuator behavior from live sensor data and adapt the climate automatically. Vertical farming, which stacks crops in layers to squeeze productivity from land-scarce cities, has embraced the digital twin approach, with one framework using a genetic algorithm to continuously retune RGB LED lighting against measured plant performance; the adaptive strategy consistently beat static configurations. Hydroponics, the soilless cultivation of roots bathed in nutrient solution, offers up to 90 percent water savings through closed-loop recycling, and machine learning and deep learning models now predict lettuce growth, flag abnormal plant conditions, and, with explainable AI, optimize even Thai basil production.

Water may be AI’s most dramatic success story. As climate change tightens the screws on global freshwater supplies, the review found smart irrigation systems consistently outperforming manual scheduling. An Arduino-based autonomous irrigation rig pairing soil-moisture and air-humidity sensors with an adaptive machine-learning controller achieved a water-use efficiency of 87.97 percent alongside higher plant survival, more uniform growth, and better visual quality than conventional watering. An AI-integrated framework combining IoT sensor networks with predictive analytics reached roughly 80 percent decision accuracy and simulated water-use efficiency gains of up to 25 percent, while a K-nearest-neighbors model predicting irrigation needs hit 98.3 percent accuracy in field trials. In a cloud-deployed greenhouse system for cherry tomatoes, models spanning multiple linear regression, support vector machines, random forests, extreme gradient boosting, and recurrent architectures predicted optimal irrigation with coefficients of determination around 0.82 to 0.83, translating into 17.8 percent heavier individual fruits and a 20.7 percent improvement in water-use efficiency under real growing conditions.

Computer vision is closing in on the pathogens, pests, and post-harvest logistics that drain harvests. An enhanced Faster R-CNN with multiscale feature fusion detected strawberry diseases at 92.18 percent mean average precision, and a pruned YOLOv5s model classified melon leaf diseases at 96.7 percent mAP with real-time inference. The most striking figure belongs to a hydroponic lettuce system called CNN-WOPNet, which married a ParNet-attention-enhanced convolutional network to a Walrus Optimization Algorithm and achieved 99.54 percent classification accuracy, 99.60 percent precision, and a 99.61 percent F1-score in identifying leaf diseases under extreme environmental conditions. Robotics is following the same trajectory. A tomato-monitoring robot using RGB-D cameras, LiDAR, and Faster R-CNN located fruit at 88.6 percent accuracy even with obscured samples, a YOLOv5-powered machine graded greenhouse mushrooms by size with 96 percent accuracy, and a vision-guided sowing robot cut vegetable planting operation time by 51 percent. Meanwhile, hybrid CNN-LSTM models fusing spatial and temporal features are sharpening soybean yield forecasts built from climate indicators and hydrological records, improving logistics planning and reducing waste.

But the review refuses to celebrate uncritically, and its quality appraisal exposes a soft underbelly beneath the headline numbers. Most included studies validated their models only through train-test splits or k-fold cross-validation on datasets from a single location, season, or production system; independent external validation, the gold standard for proving that a model generalizes, was rare. That means the published accuracies may overestimate real-world applicability. The authors also flag a metrics obsession: studies routinely report accuracy, precision, recall, F1-score, intersection-over-union, mean average precision, and RMSE, while computational efficiency, inference time, robustness under field conditions, and long-term operational reliability go largely unmeasured. Only a small fraction of the systems were ever tested in genuine agricultural environments, with many remaining at the laboratory or prototype stage. The team rated the overall risk of bias as low to moderate, stemming from dataset selection, controlled validation environments, limited external testing, and a publication bias that favors high-performing models, a reminder that failure cases, which matter most to practicing farmers, rarely make it into print.

The path forward, the authors argue, runs through hybridization and hardening. Multimodal systems combining machine learning, deep learning, and IoT still represent only 5.26 percent of the literature, yet they consistently post the strongest results, and AI-enabled edge devices are already delivering real-time pest and disease detection with high accuracy, low latency, and reduced energy consumption. Scaling those wins will demand standardized benchmarks, field-scale trials across locations and seasons, energy-efficient models that run on inexpensive hardware, explainability to win farmer trust, and robust data security for farms that have become, in effect, distributed computing networks. The review also ties the technology to crop genetics: AI-driven high-throughput phenotyping using RGB, hyperspectral, thermal, fluorescence, and 3D imaging is accelerating the identification of yield, stress, and disease-resistance traits that genomic tools can then target for breeding. What emerges is a portrait of agriculture in mid-transformation, no longer purely a craft of soil and weather and not yet a fully autonomous industry, but unmistakably becoming an information science. The success of that transition will help decide who eats in 2050.

The paper rigorously examines how AI tools, including machine learning algorithms and data analytics, are being applied in horticultural settings to augment productivity while minimizing environmental impact. The authors delve into case studies showcasing successful implementations of AI technologies in crop management, pest control, and resource optimization, thereby laying a robust foundation for understanding the potential of these innovations.

One of the focal areas of the review is the role of AI in enhancing decision-making processes in horticultural practices. The utilization of predictive analytics allows farmers to achieve higher yields by enabling precise interventions based on data-driven insights. The authors discuss various AI models that can forecast weather changes, soil conditions, and pest outbreaks, empowering farmers to plan and execute their strategies more effectively. As climate change continues to pose challenges to agriculture, this aspect is particularly vital for ensuring food security.

Moreover, the systematic review underscores the importance of automation in the future of sustainable horticulture. Robotics, propelled by AI, is taking center stage in activities such as planting, watering, and harvesting. The authors spotlight several robotic systems and their integration with AI for optimal performance. This convergence not only enhances efficiency but also reduces labor costs and the physical toll on agricultural workers. With these advancements, the sector is on the brink of a significant technological renaissance.

A crucial highlight is the application of AI-driven precision agriculture techniques which aim to utilize resources judiciously. The review meticulously points out how sensor technologies and drones equipped with AI capabilities facilitate real-time monitoring of crops. This real-time data collection is critical for farmers to make informed decisions regarding irrigation and fertilization, thus conserving water and reducing chemical runoff. By minimizing the use of harmful inputs, sustainable practices can flourish, protecting ecosystems while bolstering agricultural output.

An interesting notion presented in the review is the integration of AI with Internet of Things (IoT) devices. This symbiotic relationship is reshaping the agricultural landscape by promoting smart farming solutions. The interconnectedness of devices provides a comprehensive ecosystem wherein data flows seamlessly, allowing farmers to assess metrics on soil quality, plant health, and environmental conditions instantaneously, further informing their sustainable practices.

The review also emphasizes the challenges that accompany the integration of AI in horticulture. While there is immense potential, factors like high initial costs, the need for training personnel, and infrastructural limitations pose significant hurdles. The authors discuss potential strategies for overcoming these obstacles, including governmental incentives and partnerships between private enterprises and educational institutions. Their insights are crucial for stakeholders aiming to foster a conducive environment for AI adoption in agriculture.

Furthermore, the socio-economic implications of deploying AI technologies in horticulture are examined. The review raises pertinent questions about equity and access in the farming community. As advanced tools may favor larger operations with greater resources, there is a need for frameworks that ensure smallholders and marginalized groups can also reap the benefits of AI innovations. This balance is crucial for sustaining livelihoods and promoting inclusivity in agricultural advancements.

Another aspect explored involves the ethical considerations surrounding AI in horticulture. The authors advocate for transparent practices to prevent the misuse of data and ensure the privacy of farmers. By establishing ethical standards, the agricultural community can foster trust in AI technologies. This is particularly relevant as the sector transitions towards a more data-driven paradigm, with an increased reliance on algorithms in decision-making.

The systematic review offers a forward-looking perspective, contemplating the future landscape of horticulture influenced by AI. The authors speculate about upcoming trends, such as the increased capability of AI to analyze vast datasets to generate actionable insights, leading to highly specialized horticultural practices tailored to individual farms. This granularity is expected to enhance sustainability and productivity alike.

Additionally, the review presents a critical analysis of the current state of research and development in this domain. Highlighting gaps in literature, the authors call for interdisciplinary collaborations bringing together agronomists, data scientists, and policy-makers. Such collaborations could spawn innovative solutions and drive forward the agenda of sustainable horticultural production.

In conclusion, the systematic review by Chaudhary et al. serves as a clarion call to harness the power of AI in sustainable horticulture. As the global population continues to rise, the pressure on agricultural systems intensifies, demanding innovations that can meet these challenges head-on. This exhaustive study not only showcases the current applications of AI in horticulture but also articulates a vision for a future where technology and sustainability coalesce to create a resilient agricultural ecosystem.

It is clear that further advances in AI are imperative for overcoming existing challenges and paving the way for sustainable practices in horticulture. The authors’ insights provide a roadmap for researchers, practitioners, and policymakers to collaborate more effectively, ensuring that the agricultural sector can thrive amid uncertainty. Sustainable horticulture powered by artificial intelligence is not just an ideal but an attainable reality, promising a future where technology enriches natural ecosystems rather than detracts from them.

Subject of Research: AI in Sustainable Horticulture

Article Title: Artificial intelligence in sustainable horticultural production: a systematic review

Article References: Chaudhary, V., Kumar, M., Chaudhary, V. et al. Artificial intelligence in sustainable horticultural production: a systematic review. Discover Plants 2, 372 (2025). https://doi.org/10.1007/s44372-025-00455-2

Image Credits: AI Generated

DOI: https://doi.org/10.1007/s44372-025-00455-2

Keywords: Artificial Intelligence, Sustainable Horticulture, Precision Agriculture, Smart Farming, Robotics in Agriculture, IoT in Agriculture, Data-Driven Decision Making, Agricultural Innovation, Environmental Sustainability.