Climate Change and Sustainable Agriculture: Water, Soil, Carbon, Crops and Livelihoods in Transition

Artificial intelligence is moving horticultural automation from isolated sensing tasks towards systems that can perceive crop state, predict outcomes, optimise decisions and execute physical actions. Yet the evidence base is uneven. High performance on curated image sets is common, whereas cross-season, cross-cultivar and commercial-scale validation remains much less frequent, and the transition from prediction to safe autonomous control is still incomplete. This critical narrative review synthesises research on artificial intelligence in fruit, vegetable, ornamental and controlled-environment horticulture, with emphasis on the progression from crop monitoring and phenotyping to decision support, robotic operations and autonomous production. Literature published from 2015 to 18 July 2026 was considered, with earlier foundational studies retained where necessary. Evidence was appraised for ecological validity, external validation, sensor and dataset dependence, comparison quality, deployment realism and integration across the perception-decision-action chain. Computer vision has achieved substantial advances in disease recognition, fruit detection, crop-load estimation and maturity classification, but benchmark accuracy often overstates deployment readiness because occlusion, illumination, cultivar change and sensor drift alter the operational data distribution. Spectral sensing and multimodal approaches expand the range of measurable traits, although calibration transfer, data cost and interpretability remain constraints. Controlled-environment horticulture provides the strongest demonstrations of closed-loop artificial intelligence, including autonomous greenhouse experiments, but most reinforcement-learning and hybrid control studies still rely heavily on simulation. Robotic harvesting illustrates the central translational problem: perception can be accurate while cycle time, reachability, collision avoidance, detachment and crop damage limit whole-system performance. The evidence therefore supports a shift from model-centric optimisation to system-level validation, with shared benchmarks, uncertainty-aware multimodal models, hybrid mechanistic-learning architectures, safer human-in-the-loop control, crop-robot co-design and rigorous economic and environmental evaluation. Artificial intelligence is already valuable as an augmentative technology in horticulture; fully autonomous production will depend on reliable integration rather than continued gains in algorithmic accuracy alone.