Comparative Analysis of Statistical, Machine Learning, and Deep Learning Approaches for Frost Prediction in the Peruvian Altiplano

Frost events represent a critical climatic hazard for agricultural systems in the Peruvian highlands, impacting approximately 74% of rural communities in the Puno region. This research addresses the question of whether machine learning (ML) and deep learning (DL) approaches can significantly outperform traditional statistical methods for frost prediction in extreme high-altitude tropical conditions, achieving sufficient accuracy for operational early warning systems. We present a comprehensive evaluation of twelve forecasting models for predicting daily minimum temperatures, utilizing NASA POWER satellite data (2000-2025) from thirteen meteorological stations across the Alti-plano plateau (121,056 observations). The study implements and compares traditional statistical approaches (SARIMAX, Holt-Winters, Prophet, STL+ARIMA), machine learning algorithms (Random Forest, Support Vector Machines, XGBoost), deep neural network architectures (Multilayer Perceptron, LSTM, 1D-CNN), a hybrid SARIMA+ANN model, and an optimized ensemble approach. The ensemble model, integrating XGBoost, LSTM, and Random Forest through weighted averaging, demonstrated superior performance with RMSE=1.65°C and TSS=0.87, representing a 35% improvement over the best-performing statistical method. Individual analysis revealed XGBoost achieved RMSE=1.78°C with exceptional feature interaction modeling, while LSTM networks exhibited remarkable temporal pattern recognition with recall=0.88 for frost event detection. These findings validate the effectiveness of nonlinear approaches for operational forecasting under extreme climatic conditions and offer a robust framework for early warning systems that could substantially mitigate agricultural losses in vulnerable high-altitude communities.