Environmental and operational conditions (EOC) exert an adverse influence on monitoring systems that utilize permanently installed sensors. This proves particularly evident in the era of machine learning (ML). Since their effects can cause stronger signal changes over time than the phenomena to be detected, EOC must be adequately addressed. Emerging computer-vision (CV) use cases, such as intelligent radar units bound to the mast of wind power plants, share similar challenges with established commercial applications like static video surveillance. Those scenarios suffer from baseline observations for training that may turn out unrepresentative of deployment where such stationary camera-type devices experience temporal data shift due to their evolving non-stationary surroundings. Here, ML-based decisions become prone to false positives, resulting in high downstream investigation costs as well as low user acceptance of these smart systems. Accordingly, a rich literature exists on adapting to various types of drift. However, scenarios remain commonly neglected where (a) only noisy and potentially misleading labels are available as feedback for re-training during operation, (b) no complementary sensors are installed, and (c) changes occur not necessarily in the more obvious marginal distribution of variables, but in their joint probabilities that subtly change over time. This paper therefore focusses on solution strategies involving (i) powerful deep pre-trained models promising robustness (ii) preventive augmentation of training data, (iii) continual learning that constrains model weights over time to mitigate improper shortcut learning. Triggered by recent observations in tower-radar CV research, data drift is studied across several datasets which thereby highlights the problem's generality. A computational framework is introduced to systematically investigate the complex setup empirically. Probing different ML models, a phase-transition phenomenon in an overfitting tree classifier is incidentally discovered. The paper presents both practical recommendations for CV-based tower-radar monitoring as well as findings of interest for the broader community of researchers in ML.