Weather Integrated Forecasting of Helicoverpa armigera using ARIMA and ANN Models Trained on a Decade of Pheromone Trap Data

Technical Summary: Weather Integrated Forecasting of Helicoverpa armigera using ARIMA and ANN Models

Problem Statement
Helicoverpa armigera (Hübner) is a highly polyphagous and migratory pest causing significant yield losses in diverse cropping systems, particularly in the Tarai region of Uttarakhand, India. Its population dynamics are heavily influenced by climatic variability, including temperature, humidity, and rainfall. While Integrated Pest Management (IPM) offers a sustainable alternative to indiscriminate chemical control, its success relies on accurate, timely forecasting to trigger interventions before pest populations reach economic thresholds. Despite the critical need for predictive capabilities in the humid subtropical agro-climatic conditions of Pantnagar, systematic long-term forecasting efforts utilizing advanced time-series and machine learning approaches have been limited.

Methodology
The study developed and compared two distinct modeling frameworks to forecast adult H. armigera populations using a decade (2015–2025) of weekly pheromone trap data and associated weather variables (mean temperature, relative humidity, rainfall, and sunshine hours).

1. Data Preprocessing: Weekly time-series data were inspected for trends and anomalies. Outliers were corrected via imputation or removal. Stationarity was assessed using the Augmented Dickey–Fuller (ADF) test, with differencing applied where necessary for ARIMA modeling. For Artificial Neural Network (ANN) analysis, input features were normalized using min–max scaling.
2. Modeling Approaches:
   • ARIMA: An Autoregressive Integrated Moving Average model was fitted to capture linear temporal dependencies and seasonality. Model selection was based on minimizing the Akaike Information Criterion (AIC) and Bayesian Information Criterion (BIC). The optimal configuration identified was ARIMA (1, 0, 1) (1, 0, 2).
   • ANN: A feed-forward multilayer perceptron was trained to capture nonlinear relationships between lagged pest population data and weather predictors. The optimal architecture identified was 7-26-1 (7 input neurons, 26 hidden neurons, 1 output neuron), utilizing ReLU activation in hidden layers and a linear activation in the output layer. Training employed the Adam optimizer with early stopping to prevent overfitting.
3. Validation: Data were partitioned into 80% for training and 20% for testing, preserving chronological order. Model performance was evaluated using Mean Absolute Percentage Error (MAPE), Root Mean Squared Error (RMSE), Mean Absolute Error (MAE), Mean Absolute Scaled Error (MASE), and the Coefficient of Determination ($R^2$). Residual diagnostics (ACF plots, QQ plots) were conducted to ensure residuals approximated white noise.

Key Results

• Model Performance: The ARIMA (1, 0, 1) (1, 0, 2) model exhibited the highest overall predictive reliability with the highest $R^2$ (0.76) and lowest MAPE (107.91), though it had a higher RMSE (6.62). In contrast, the ANN (7-26-1) model achieved a lower $R^2$ (0.644) but demonstrated superior short-term predictive accuracy for specific metrics, yielding lower RMSE (3.63) and MAE (2.144) compared to ARIMA (RMSE = 6.62, MAE = 2.67).
• Behavioral Differences:
  • The ARIMA model effectively characterized long-term seasonal trends and provided a smooth representation of interannual variations. However, it tended to underestimate abrupt population peaks and exhibited wider confidence intervals during outbreak periods.
  • The ANN model successfully captured nonlinear fluctuations and sharp population peaks, demonstrating higher sensitivity to complex interactions between weather variables and pest activity. While it occasionally overestimated peak magnitudes, it tracked observed trends more closely during high-incidence periods, resulting in lower error magnitudes (RMSE and MAE) for short-term forecasting despite a lower overall $R^2$.
• Forecasting: Forecasts for 2025–2027 indicated recurring seasonal patterns consistent with the multivoltine life cycle of H. armigera, with peaks predicted during spring (March–April) and post-monsoon (October–November). The ANN model generated more dynamic predictions responsive to cyclic environmental inputs, while ARIMA provided conservative, stable estimates.

Significance and Claims
The authors assert that while ARIMA models remain valuable for interpreting underlying periodicity and long-term trends (evidenced by their higher $R^2$), ANN models offer enhanced precision and adaptability for real-time, short-term forecasting of H. armigera in the Pantnagar region, as indicated by their lower RMSE and MAE. The study highlights the value of ANN models as practical early warning systems within IPM programs for capturing nonlinear population dynamics. By integrating weather-based forecasting models into pest surveillance, agricultural systems can transition from reactive to preventive management strategies. This approach aims to optimize the timing of control measures, reduce unnecessary pesticide applications, and minimize crop losses, thereby supporting sustainable crop protection and food security in northern India. The research fills a gap in quantitative forecasting frameworks for this specific agro-climatic zone, offering a reference for improving precision pest management under variable climatic regimes.