Improved sap flow prediction: A comparative deep learning study based on LSTM, BiLSTM, LRCN, and GRU with stem diameter data
Introduction Recording sap flow in plants is essential to understanding water usage, especially for herbaceous species like tomatoes. While plant physiology research has progressed, there remains a gap in applying sap flow sensor data to these species. In this study, the predictive capabilities of R...
Saved in:
| Main Authors: | , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
Elsevier
2025-12-01
|
| Series: | Smart Agricultural Technology |
| Subjects: | |
| Online Access: | http://www.sciencedirect.com/science/article/pii/S2772375525003387 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Introduction Recording sap flow in plants is essential to understanding water usage, especially for herbaceous species like tomatoes. While plant physiology research has progressed, there remains a gap in applying sap flow sensor data to these species. In this study, the predictive capabilities of Recurrent Neural Network (RNN) architectures—LSTM, GRU, BiLSTM, and LRCN—are explored for sap flow estimation in tomato plants using stem diameter variations as the sole input. Unlike existing studies that rely on multi-variable environmental data from large-scale datasets such as SAPFLUXNET, COCO or KAGGLE this research is based on in-house experimental data collected in close collaboration with a sensor developer and farmers. The experimental setup reflects practical conditions relevant to controlled environment agriculture. To the best of the authors’ knowledge, this is the first study to investigate the potential of RRN deep learning models to infer sap flow directly from stem diameter signals in tomato plants. A comprehensive performance comparison of the models is presented under varying input time windows, with a discussion on implications for real-time irrigation and plant monitoring solutions. Deep learning models were designed using four advanced Recurrent Neural Networks (RNN) architectures: LSTM, BiLSTM, LRCN, and GRU, trained with past sap flow and stem diameter data from tomato plants. Based on the last three hours of data, the models predicted sap flow for the next hour. Metrics like Root Mean Squared Error (RMSE), Mean Absolute Error (MAE), and R² were used to evaluate performance, and early stopping was applied to prevent overfitting during training. The LSTM model achieved the lowest RMSE, excelling at short-term sap flow prediction. However, both BiLSTM and GRU models performed well overall, particularly in capturing, more significant fluctuations and peaks. R2 0.83 values across all models were around 7.2, with MAE values below 5.8, demonstrating robust predictive potential. These results suggest that advanced deep learning models, particularly BiLSTM, can significantly improve the prediction of plant sap flow, enhancing efficiency in water management in precision agriculture. Future research could apply these models to other herbaceous species. |
|---|---|
| ISSN: | 2772-3755 |