Evaluation of Test-Time Adaptation for Computer Agriculture Crop-Disease Classifiers Under Seasonal Camera Sensor Drift: An Artificial Intelligence Entropy-Minimization Bound for Misdiagnosis Risk

Crop-disease classifiers deployed on farm cameras face a recurring obstacle: the colour and texture statistics of the images they receive shift with the season, the light and the sensor, while the classifier's parameters remain frozen at training time. Test-time adaptation (TTA) responds by adjusting the model on unlabelled field images as they arrive, and entropy minimisation is the most widely used objective for doing so. the present analysis asks what guarantee entropy-minimising TTA can actually offer against misdiagnosis when the drift is a slow, seasonal displacement of the input distribution rather than a sudden corruption. The work is theoretical and methodological: no field data are analysed. A bound is derived that relates the target-domain misdiagnosis risk of an entropy-adapted classifier to three quantities — the source-domain risk, the divergence between the source and target feature distributions, and the sharpness of the entropy minimum the adaptation reaches. The bound shows that entropy minimisation controls only the third term, and that its benefit is capped by the first two; under a seasonal drift that moves decision boundaries rather than merely blurring them, the bound can be loose enough that adaptation increases risk. The result is used to specify the conditions under which TTA should be evaluated before deployment, and to argue that seasonal drift in crop-disease imaging requires an evaluation protocol that reports the divergence term alongside accuracy.