Implementing of Pest Forecasting Analytical System: Quality Assurance Testing

This paper describes a season-aware automated statistical testing approach for quality assurance of machine learning systems predicting pest numbers on cereal crops. An automated testing framework was developed, specifically tailored to long-term agricultural time series. It focuses on the neural network system designed to predict pest population dynamics in Kazakhstan using a Transformer Encoder architecture trained on real monitoring and meteorological data collected between 2005 and 2022. We introduce domain-informed stress-testing scenarios – including noise injection, temporal shifts, missing data, and controlled perturbations of weather-related inputs – and integrate them directly into the machine learning pipeline to enable regular reliability assessment when new data become available. Unlike existing quality assurance methodologies such as ML Test Score or DeepXplore, our approach is explicitly adapted to the challenges of seasonal and noisy agricultural data and provides automated reliability reporting. The proposed approach increases model transparency, reduces deployment risks, and supports iterative improvements by embedding testing within the machine learning pipeline. This ensures the system's adaptability to seasonal variations, resilience to noisy and incomplete data, and readiness for integration into decision-making processes in precision agriculture.