Safety-Aware Cascaded Inference for Crop Damage Assessmentwith Controlled Error Trade-offs

In picture-based agricultural insurance for smallholder farmers, missed damagedetections carry substantially higher cost than false alarms: a farmer whosustained real losses receives no payout, while unnecessary expert review isoperationally costly but reversible. Standard multi-class classifiers optimizeglobal accuracy but provide no mechanism to operationalize or control thisasymmetric cost structure at inference time.We propose CascadeCropNet, a two-stage cascade architecture that canbe calibrated to satisfy a target recall constraint (Rec-Damaged ≥ 0.95)under a chosen operating point through threshold selection. A lightweightSentinel model performs binary health triage; samples exceeding a calibrateddamage probability threshold τ are escalated to a specialist Expert model forfine-grained diagnosis. This design provides explicit, deployment-time controlover the safety–efficiency trade-off without retraining.Evaluated on the Eyes on the Ground dataset (28,077 images, 23,804retained after label consolidation, from Kenyan smallholder maize farms),the cascade achieves Rec-Damaged = 0.974 at τ = 0.5, reducing missed damage cases by up to 54% at a selected operating point relative to a flat baseline. Under evaluation alignment, the representational gap reduces to+0.008 F1-macro, confirming that the contribution is architectural ratherthan representational. Under input degradation, the system prioritizes esca-lation over confident misclassification, reflecting error containment througharchitectural isolation rather than intrinsic model robustness.These results demonstrate that cascade architectures can operationalizeand control safety-oriented decision constraints through calibrated routing indeployment settings where reliability and controllability matter more thanaggregate accuracy. These properties depend on threshold calibration anddeployment conditions and do not constitute guarantees under arbitrarydistribution shift.