Validation of AI models for ITCZ Detection from Climate Data.

The Climate Change topic itself and associated life adjustments became a top global problem of the XXI century. It requires immediate attention and long-term solutions. Physical tracing of weather changes, performed without Artificial Intelligence (AI), is not capable of detection in a timely manner and correctly classifying dangerous situations, which is crucial for making rapid decisions and taking immediate actions on the ground. AI and Machine Learning (ML) systems are currently dominating solutions in Computer Science research fields and the issues of their testing and validation remain a critical open problem due to their uncertain outcomes. We use test automation to validate and assure the quality of various AI systems for one kind of weather prediction. The subject of our study is Inter-Tropical Convergence Zones (ITCZs). ITCZs play an important role in the global circulation system and even small changes in their patterns can cause severe droughts or flooding as well as other disasters like hurricanes. Global warming is causing more ITCZ scenarios as shown in weather data, which makes physical detection infeasible. Our research initially discovered that ITCZ detection based on a physical model alone could misclassify some unexpected situations when the double bands occur at different places with similar intensity or at the same places with different intensities. We then designed experiments to train AI models to detect ITCZs with test automation to collect results. We further trained several AI models with focus on VGG-16, VGG-19, Xception and MobileNETV2 models, collected and compared their results through test automation. Our exhaustive trials eventually achieved a 96.8% accuracy, which might be the best AI model to detect ITCZs with test automation and without human intervention. These results show that test automation can contribute to the selection of optimum AI models.