Learning-Rate-Free Adaptive Optimization for Artificial Intelligence Agriculture Crop Disease Classifiers Under Nonstationary Field Image Streams: A Computer Convergence-Regret Evaluation

Crop disease classifiers deployed on field image streams face a learning-rate dilemma: fixed step sizes either track nonstationary disease presentation too slowly or oscillate under label noise, and no single rate is optimal across the seasonal and illumination shifts that characterise agricultural image capture. The framework couples a coin-betting online convex optimization layer with a per-parameter adaptive preconditioner, removing the need for a tuned base learning rate while retaining the diagonal scaling that stabilises deep classifier training. Convergence is established for the average regret of the composite update under bounded gradient and bounded drift assumptions, and a high-probability bound is derived for the nonstationary regret that scales with the cumulative drift rather than the horizon. A replication-based evaluation was conducted across 2000 design cells spanning cluster counts of 8, 16, 24 and 32 and effect sizes of 3, 5 and 8 percentage points, with an intraclass correlation of 0.64, between-cluster standard deviation of 6.0 and residual standard deviation of 4.5; 54 clusters were required for 80% power at a 5 percentage point effect. The analysis characterises the operating characteristics of the design under the declared variance structure and does not use observational data. The framework provides a principled route to removing learning-rate tuning from the deployment pipeline of agricultural disease classifiers, and the evaluation clarifies the cluster counts at which such classifiers can be reliably compared.