• This study applies a pipeline-centric lens to AI early warning for pests, diseases, and drought. • It synthesizes 72 studies on data fusion, lead times, baselines, calibration, and external validation. • It distills operational guidance on multimodal fusion, physics-aware models, MLOps, and FAIR/OGC interoperability. • It offers a decision-ready playbook linking calibrated probabilities to alert tiers and cost–loss actions. Climate-amplified droughts and rapidly spreading pests and plant diseases are eroding agricultural productivity and farmer livelihoods. Advances in Earth observation, in-field sensing, and machine learning (ML) promise AI-enabled early warning systems (EWS), yet operational guidance remains fragmented across hazards and technology layers. This study maps how predictive pipelines for agricultural pests, plant diseases, and drought are designed, validated, and operationalized—and to identify gaps that constrain real-world, people-centred EWS. We conducted a PRISMA-ScR scoping review (January 1, 2000–September 30, 2025) across the Web of Science, Scopus, and Google Scholar databases, utilizing dual screening, dual extraction, and a pipeline-centered charting template. From 1,842 records, 72 studies met the inclusion criteria. Pipelines predominantly fused earth observation (optical/SAR) with reanalysis/forecast data, while in-situ IoT was chiefly used in field-scale pest/disease applications. Forecast horizons concentrated at 1–4 weeks (pests/diseases) and 1–6 months (drought). Baseline comparators were reported in 57% (persistence) and 39% (climatology) of studies; probabilistic outputs appeared in 38%, with calibration evidence in 21% and cross-region external tests in 24%. Operational features included versioning/registries (46%), drift detection (29%), scheduled retraining (39%), multilingual/offline delivery (32%), and interoperability claims aligned with FAIR principles (Findable, Accessible, Interoperable, Reusable) and OGC (Open Geospatial Consortium) geospatial standards/APIs (25%) (18/72 studies). Pipelines performed best when multimodal, phenology-aware fusion incorporated relevant teleconnections; when model inductive bias reflected hazard physics (graph diffusion; physics-informed constraints); and when verification was decision-grade (skill vs. baselines under spatial/temporal blocks, with calibrated probabilities linked to alert tiers and cost–loss framing). Gaps persisted in external validation, reliability reporting, interoperable artifacts, and equity-by-design, and only a minority of studies assessed downstream changes in farmer behavior, avoided loss, or livelihood outcomes. A coherent, pipeline-engineered approach can move AI EWS from prototype to people-centred scale, but stronger evidence on actual livelihood impacts is still needed.
Md Nazirul Islam Sarker (Sun,) studied this question.