FICformer: A Hybrid Fuzzy-Transformer Framework for Agricultural Data Imputation and Forecasting

Abstract

Smart agricultural developments are moving towards intelligent sensor networks that are now gathering realtime environmental information, including temperature, humidity, and CO2 concentration. Even though partial sensor measurements can often decrease model performance, such real-time data streams enable predictive analytics to enhance crop development, and require effective data imputation and forecasting methods. A Spark object is employed to process the SF24 data to handle large volumes of data and perform remote computations, where the data are multivariate time-series readings of agricultural sensors. Irregular crop growth forecasting is caused by the fact that conventional LSTM algorithms are often unable to deal with anomalies in data. The FICformer technique combines a transformer-based encoder– decoder design with fuzzy Bayesian imputation to overcome these drawbacks. To provide data completeness and reliability, imputation component employs Bayesian-based fuzzy guesses to reconstruct missing sensor data. A pooling layer minimizes redundancy and computing time and a dimensional temporal attention approach is employed to obtain temporal dependencies and inter-variable correlations. GRU and BidirectionalGRU were applied to FICformer to design hybrid and layered designs that enhanced performance. The Stacked Former configuration had the highest predictive accuracy of smart agriculture systems with an RMSE of 2.746549, compared to baseline LSTM, FICformer, and hybrid approaches, with outstanding predictive accuracy.