Predictive Crop Growth Modeling Using Transformer‑Enhanced Imputation of Agricultural Sensor Data

Abstract

Smart farmland is getting better with the help of smart sensor networks that collect real-time data about the environment, like CO2 levels, temperature, and humidity. These continuous data streams allow predictive analytics to help crops grow better, but sensor readings that aren't full often make models less accurate, so they need effective ways to fill in missing data and make predictions. A Spark object is used to handle big amounts of data and make distributed computation easier for the Smart Farming Data 2024 (SF24) dataset, which is made up of multivariate time-series measurements from agricultural sensors. Conventional Long Short-Term Memory (LSTM) models often have trouble with missing or corrupted data, which makes it hard to predict crop growth accurately. To get around these problems, the FICformer method combines fuzzy Bayesian imputation with an encoder–decoder architecture based on transformers. The imputation part fills in missing sensor data using Bayesian calculations and fuzzy estimates to make sure the data is full and accurate. A dimensional temporal attention method is used to find correlations and dependencies between variables over time, and a pooling layer cuts down on duplicate data and computation time. To improve efficiency, hybrid and stacked configurations were made by putting together FICformer with GRU and Bidirectional GRU layers. The Stacked Former configuration got an RMSE of 2.746549%, which shows that it was very good at making predictions and did better than baseline LSTM, FICformer, and hybrid methods for smart agriculture systems