Forecasting soil moisture and pH on edge for shallot cultivation using temporal fusion transformer

This paper presents an edge-enabled forecasting framework for soil moisture and pH in shallot cultivation using a modified Temporal Fusion Transformer (TFT). Instead of the conventional LSTM-based encoder decoder, we incorporate a cross-attention block to better align historical observations with known-future inputs while maintaining interpretability through attention analysis. To reduce high-frequency noise from the IoT sensing network, Savitzky–Golay (SG) filtering is applied during preprocessing. Field data were collected from 300 Brebes shallot plants over 74 days under three soil-moisture regimes (Ideal, Wet, and Dry). Model training and inference were performed on an NVIDIA Jetson Orin edge device to support real-time operation. Performance was evaluated using MAE, RMSE, quantile loss, and prediction-interval coverage, with Elastic Net and weight decay regularization. Ablation results show that cross-attention improves forecasting consistency, reducing MSE/RMSE to 15.51/3.69 compared with 16.93/3.83 when removed, while static enrichment stabilizes variability across sensing points. SG filtering further reduces variance and pinball loss, contributing to more stable edge-side processing. On Jetson Orin, the SG-enabled pipeline achieved 0.12 ms/sample latency and 8,505 samples/s throughput; although latency differences are small, throughput gains are more pronounced relative to the unfiltered setup. The system supports bidirectional 4G telemetry for probabilistic forecasting and web-based monitoring. This study is limited to a single soil type (Inceptisol) and one agro-climatic site; extending validation across different soil textures and regions remains an important next step.