Smart agriculture is booming, but disconnected AI tools are holding farms back

A new paper argues that AI has already improved agricultural perception, prediction and task-level decision-making, but the next major challenge is to move from isolated smart tools to integrated, field-ready farming systems.

Published in Sustainability, the study "Artificial Intelligence in Smart Agriculture Across the Production-to-Postharvest Continuum: Progress, Challenges, and Future Directions" reviews recent AI applications across the tillage, sowing, management and harvesting workflow, while also extending the discussion to postharvest handling, with a focus on how smart agriculture can move from stage-specific automation toward system-level autonomy.

AI is moving smart farming beyond isolated tools

Smart agriculture is a response to mounting pressure on global food systems, including population growth, labor shortages, resource scarcity and climate instability. Agriculture can no longer rely only on mechanization or manual inspection to meet these challenges. Modern farming increasingly requires systems that can sense changing field conditions, interpret complex data, make timely decisions and guide machinery or robotic actions with limited human intervention.

AI is an umbrella technology that includes machine learning, deep learning, computer vision, robotic control, reinforcement learning, hyperspectral imaging and decision-support algorithms. These tools are being used across the farming cycle to improve soil analysis, guide tillage depth, monitor seed placement, detect weeds and diseases, regulate irrigation and fertilizers, estimate crop growth, support harvesting robots and assess product quality after harvest.

The study gives a full-chain view of agricultural AI. While many previous reviews have focused on single areas such as crop monitoring, irrigation, harvesting robots or plant disease detection, this review organizes the field around the full production-to-postharvest continuum, showing that the major problem is no longer whether AI can perform individual farm tasks. The bigger issue is whether these tasks can be linked into a coherent smart farming system.

Farm operations are tightly connected. For instance, soil preparation affects seed placement while sowing quality affects crop emergence. Crop emergence shapes management needs for irrigation, fertilization and plant protection and management decisions influence harvest timing and produce quality. Harvest outputs then feed into grading, storage, logistics and traceability. If AI tools work only at one stage, their value remains limited.

According to the review, AI applications in tillage and field preparation are shifting from simple sensing toward integrated perception and control. Soil-property prediction increasingly uses spectroscopy and machine learning to estimate conditions without relying only on laboratory tests. AI models can interpret spectral data to predict soil properties and support decisions on tillage, salinity estimation and field readiness. Intelligent tillage machinery is also using IoT monitoring, edge computing, engine-load sensing and control algorithms to regulate tillage depth, reduce traction resistance and improve energy efficiency.

Field path planning is another critical area. AI and optimization algorithms are being used to reduce nonworking travel, cut fuel consumption, minimize headland operations and improve route efficiency for tractors, harvesters and drones. In more complex terrain, algorithms can support route planning across irregular plots or hilly areas, where traditional coverage methods struggle. These developments suggest that AI is not only improving measurement, but also helping machines act more efficiently in the field.

However, the review stresses that tillage-stage AI still faces transferability problems. Models trained on one soil type or terrain may not work well elsewhere. Field conditions change through moisture, compaction, slope, residues and weather. Sensing accuracy alone is not enough unless AI outputs can be connected to machinery control and then passed forward into sowing decisions. This is where current systems remain fragmented.

The sowing stage shows the same pattern of progress and constraint. AI-enabled precision sowing is moving toward closed-loop systems that connect seed recognition, equipment regulation and environmental decision-making. Machine vision and improved object-detection models are being used to detect small seeds, identify missed or repeated seeding and monitor seed flow under high-speed conditions. Some systems have reduced sowing miss-detection rates to below 3% under controlled or bench-top conditions.

Precision sowing is is vitally important because seed placement affects crop emergence uniformity, later field management and final yield. AI can help monitor sowing quality and support real-time adjustment of parameters such as negative pressure in pneumatic seed-metering systems. It can also support seed phenotyping and breeding by using image-based analysis to estimate seed vigor, classify seed quality and predict germination.

Still, field deployment remains difficult. Dust, vibration, poor lighting, seed-size variation and hardware limits can reduce the reliability of vision systems. A model that performs well on a bench may face errors when mounted on fast-moving equipment in dusty, uneven conditions. This creates a familiar gap across smart agriculture: laboratory accuracy is improving faster than field robustness.

Crop management is becoming more data-driven, but decision-making remains fragmented

AI applications in crop management are among the most visible areas of smart agriculture. The review identifies major progress in plant protection monitoring, water and fertilizer regulation, agrochemical application and crop growth assessment. These tasks involve more dynamic conditions than tillage or sowing because crops change over time and interact with weather, pests, diseases, soil moisture and nutrient availability.

Plant protection is moving from manual scouting toward automatic detection of weeds, pests and disease symptoms. Deep learning models can process crop images, leaf images, UAV imagery and multispectral data to classify plant diseases, identify weeds and segment crop regions. These systems can support targeted spraying, reduce pesticide waste and help farmers respond earlier to threats.

AI-based disease detection has become increasingly precise, with convolutional neural networks and attention-based models able to detect visual symptoms, classify severity and distinguish nutritional deficiencies from disease. Weed detection is also advancing through UAV imagery, deep learning and segmentation models. In some cases, AI can generate weed maps that support precision herbicide application, allowing farmers to spray only where needed rather than treating entire fields.

This shift could significantly shape sustainability outcomes. If AI can guide water, fertilizer and agrochemical use more precisely, farms can reduce waste, lower environmental pressure and improve input efficiency. Smart sprayers, for example, can use sensors and AI to assess canopy structure and adjust spray output in real time. These systems can reduce drift and improve coverage while cutting chemical use.

Irrigation and fertilization are also becoming more intelligent. AI models can process soil moisture, temperature, humidity, crop images, yield maps, climate records and crop-state data to support decisions on when and how much water or fertilizer to apply. Reinforcement learning is emerging as a method for sequential decision-making, where systems learn how to balance crop needs, weather uncertainty and resource efficiency over time.

The review points out that water, fertilizer and agrochemical decisions are not isolated. Irrigation status, nutrient demand, canopy health and pest pressure interact in real fields. AI systems that treat each input separately may miss these connections. The practical value of AI lies in converting diagnosis into action, meaning the system should not only detect stress but recommend or execute the right intervention under field constraints.

Crop growth monitoring provides another key function. AI models, UAV imaging, hyperspectral sensing and deep multitask learning can estimate leaf nitrogen content, biomass, leaf area index, chlorophyll levels and crop health. These indicators help farmers assess growth status and make decisions on topdressing, irrigation and harvest planning. Deep learning can also support yield prediction by capturing complex relationships among crop traits, weather conditions and remote-sensing data.

Despite this progress, the review highlights several weaknesses. Crop management models often struggle across crops, seasons and regions. Data heterogeneity remains a major barrier because field images, sensor readings and crop responses vary sharply across environments. Annotation costs are also high, especially for disease, weed and phenotyping datasets that require expert labeling. Limited interpretability creates another problem: farmers and agronomists need to understand why an AI system recommends an action, especially when the decision affects costs, yield and environmental outcomes.

The review also raises the issue of deployment complexity. Many AI models are computationally demanding, while farms often need real-time or near-real-time decisions on edge devices with limited power and connectivity. Cloud-based systems can support training and long-term analysis, but field equipment may require lightweight models that run locally. This makes cloud-edge collaboration a key future direction.

A deeper challenge is that management-stage AI often remains disconnected from earlier and later stages. Soil maps generated during tillage are not always used to guide irrigation and fertilization. Sowing-quality data are not always linked to crop-emergence monitoring. Harvest outcomes are rarely fed back into management models.  

Harvesting and postharvest AI expose the gap between lab success and field-scale autonomy

AI is also advancing in harvesting and postharvest handling, where the focus shifts from diagnosis and crop regulation to robotic perception, manipulation, coordination, grading and traceability. This stage is critical because labor shortages are especially severe in harvesting, and postharvest quality determines market value.

Harvesting robots rely on AI to detect fruit, assess maturity, estimate position, plan grasping and control robotic arms in cluttered environments. Computer vision, RGB-D sensing, hyperspectral imaging and geometric modeling are being used to detect fruits and guide manipulation. AI can also evaluate maturity through multimodal sensing, helping reduce the risk of harvesting crops too early or too late.

The review notes that visual recognition has improved substantially, but reliable field execution remains difficult. Robots must work under occlusion, variable lighting, overlapping leaves, irregular fruit shapes, fragile produce and moving branches. Detecting a fruit is easier than picking it safely at commercial speed. Manipulation, grasp planning and obstacle avoidance remain major bottlenecks.

Economic barriers are also significant. Fully autonomous harvesting platforms remain expensive for many small and medium-sized farms. Return-on-investment timelines may be too long for growers unless systems become cheaper, faster and more reliable. Soft-fruit harvesting is especially difficult because robots must be gentle enough to avoid damage while fast enough to compete with human labor. These conditions show that commercial readiness is not determined only by model accuracy.

Harvest coordination and route optimization are also becoming more data-driven. AI can use remote sensing, UAV imagery and semantic segmentation to map crop distribution, estimate harvest capacity and support machinery dispatch. Orchard-tree mapping and crop-boundary extraction can help create digital field maps, plan routes and estimate workload. These functions could improve coordination among harvesters, transport vehicles and storage systems.

Postharvest handling is another major area of AI growth. Image analysis, hyperspectral sensing and machine learning can support quality grading, maturity assessment, defect detection, shelf-life prediction and cold-chain monitoring. These systems can reduce subjective manual inspection and create more consistent grading standards. AI can also connect quality data with traceability systems, allowing product condition to be tracked through storage, transport and market distribution.

The review identifies traceability as a crucial future area. Once quality information is digitized, it can support batch classification, storage decisions, logistics planning and market-side transparency. AI-enabled grading, combined with blockchain or other traceability platforms, could help downstream actors anticipate product quality, manage inventories and reduce postharvest losses. But the paper warns that field-side quality data are not yet well integrated with postharvest management systems. Data standards, software compatibility and platform interoperability remain weak.

Across the full production-to-postharvest chain, the review identifies one central pattern: AI performs best when the task is clearly defined, data are controlled and outputs are local. It performs less reliably when the environment is messy, decisions are sequential and systems must communicate across stages. This means smart agriculture has made strong progress in task-level intelligence but has not yet achieved full-chain autonomy.

Last but not least, the review calls on researchers to prioritize system-level robustness, interoperability and adaptability. The authors identify four key future directions.

1. Agriculture needs multimodal foundation models that can handle different crops, environments and sensing systems with less labeled data.
2. Field deployment requires lightweight and uncertainty-aware architectures that can run on edge devices and signal when predictions are unreliable.
3. The sector needs standardized datasets and evaluation protocols that measure not only accuracy but also robustness, latency, annotation cost and deployment conditions.
4. Smart farming requires full-chain decision architectures that link sensing, diagnosis, control and feedback across the production workflow.

The review also proposes a phased roadmap. In the short term, over one to three years, progress should focus on lightweight models, uncertainty-aware systems and cloud-edge collaboration. In the medium term, over three to five years, the priority should be standardized agricultural datasets and interoperable evaluation protocols. Over the longer term, the field should move toward multimodal agricultural foundation models and full-chain digital twins that combine crop growth models with real-time sensor streams.

There are some limitations, as well. The reviewed studies use different crops, sensors, metrics and task definitions, making formal comparison difficult. It also says English-language publication limits may omit regional advances, while fast-moving AI architectures released late in the review period may be underrepresented. The tillage-sowing-management-harvesting framework is useful for organizing the field, but it may understate cross-cutting issues such as cybersecurity, data privacy and socio-economic barriers to adoption.