Intelligent monitoring and prediction system for agroclimatic variables using WSN, SARIMA, and artificial intelligence in potato crops in La Calera, Colombia.

A hybrid agroclimatic prediction system is presented, based on SARIMA time series and LSTM neural networks, trained with simulated data every 10 minutes over 3 months to detect frost events in La Calera, Colombia. The SARIMA model was obtained through the Box–Jenkins approach, applying stationarity tests, seasonal differencing, and optimal parameter search using grid search and AIC/BIC criteria. Subsequently, SARIMA predictions were integrated as an additional input into a multivariate LSTM network, trained with sliding windows of 7 days and regularization techniques (Dropout, early stopping). The LSTM architecture includes two sequential layers and was optimized with Adam. The weighted fusion (60% LSTM, 40% SARIMA) improved the accuracy of event classification. The model achieved 95% accuracy and 92.3% recall, validating its applicability for precision agriculture in frost-prone areas.