Interpretable Machine Learning for Sugarcane Harvester Performance: A Comparison of Additive and Tree-Based Models on Telematics Data

Sugarcane harvester performance varies substantially with field geometry, crop, and operator factors, yet separating these sources from telematics data while preserving engineering interpretability remains a methodological gap. This study models field efficiency (Eff) and harvesting capacity (Ca) separately from JDLink telematics, aligning model structure with each target’s response behavior. Operational data covered 105 plots across four seasons (2019/20–2022/23) from three John Deere CH570 chopper harvesters in eastern Thailand. Six engineering-relevant predictors were retained after multicollinearity screening, and linear (MLR), additive nonlinear (GAM), and tree-based models were compared under 5-fold grouped cross-validation by BaseField (87 groups). Eff was assigned to GAM (R2CV = 0.621 ± 0.114) on the basis of its threshold-like response to turning frequency; Ca was retained for MLR (R2CV = 0.681 ± 0.121), with GAM essentially tied. Train–validation gaps were substantially smaller for additive models (0.096–0.118) than for tuned tree-based candidates (GBR 0.210–0.302, RF 0.322–0.358). Turning frequency (TF) and perimeter-to-area ratio (PAR) were the strongest predictors, and a constant-turn-time partial-out test indicated that TF’s univariate effect on Eff is largely mediated by the time-budget identity. Tactical interventions (path planning, operator training, machine–field allocation) are immediately feasible, although strategic field-layout change remains constrained by smallholder land tenure.