Autonomous Irrigation under Noisy Sensor-Actuator Fusion

Short-horizon soil-moisture forecasts are potentially useful for irrigation control, but field records often combine heterogeneous signals whose timestamps, units, and physical correspondence are uncertain. We study conditional variational forecasting under this constraint using a structured conditioning vector with environmental–crop context (64 dimensions), control history (32), and plan/phenology information (8). A data audit of the supplied aligned table identifies 3,000 complete paired rows, a soil-moisture target range of 120–700 sensor units, extreme light-intensity observations, and weak cross-stream correlations. In the reported conditioning ablation, context-only conditioning gives the lowest one-step error, whereas naive 104-dimensional fusion performs substantially worse. A reported offset-and-denoising rerun reduces full-fusion MAE by 33.7% relative to naive fusion, but does not surpass context-only conditioning. We formalize why this negative result is informative: auxiliary variables cannot increase population Bayes risk under unrestricted prediction, so empirical degradation indicates missynchronization, finite-sample error, noise, optimization, regularization, or model mismatch. We therefore position the contribution as a synchronization-aware evaluation protocol rather than a novel CVAE. The study also specifies calibration, robustness, and irrigation-utility evaluations required before claims of uncertainty-aware autonomous control can be made.