Handling Sensor Faults in Hydroponics: A Deep Learning Imputation Technique for Accurate Tomato Yield Prediction

Sensor faults in hydroponic systems pose significant challenges for precision agriculture by compromising the nutrient monitoring accuracy and yield prediction reliability. Current imputation methods lack domain-specific agricultural pattern-preservation capabilities. This paper presents a novel Deep Learning Precision Imputation Model (DLPIM) based on a generative adversarial network (GAN) architecture to recover missing agricultural data in IoT-based hydroponic monitoring systems. DLPIM introduces a feature encoder-based generator with a 4-headed self-attention mechanism and a Crop Growth Rate (CGR) guided temporal processor to capture complex time-series patterns while maintaining physiological consistency. The discriminator implements dual validity and feature matching pathways, enhanced by minibatch discrimination and spectral normalization. A dataset comprising 1092 data points related to tomato growth and nutrient analysis was utilized to develop and evaluate the proposed model, demonstrating the superior performance of DLPIM compared to nine state-of-the-art imputation methods across different missing data scenarios (10-60%). The proposed model achieved optimal accuracy (MSE =2.367-2.664, MAE =0.795-0.851, R ${}^{2} = 0.990$ -0.992), while maintaining consistent performance across all missing rates. Ablation studies confirmed the effectiveness of architectural innovations, with CGR removal resulting in the most substantial performance decline (R ${}^{2} = 0.851$ at a 60% missing rate). The DLPIM framework establishes a new benchmark for agricultural time series imputation, enabling robust decision-making in precision agriculture.