Early Forecasting of Rice Blast Disease Using Long Short-Term Memory Recurrent Neural Networks

Yangseon Kim, Jae-Hwan Roh, Ha Young KimSustainability 2017 · DOI

TL;DR

This paper introduces a novel approach for predicting rice blast disease outbreaks a year in advance by leveraging Long Short-Term Memory (LSTM) recurrent neural networks. By integrating historical rice blast incidence data with climatic factors, the model achieves promising prediction accuracies, offering a valuable tool for proactive disease management and crop breeding.

What Problem Does This Paper Address?

Rice blast disease, caused by the fungus Magnaporthe oryzae, is a major threat to global rice production, leading to significant food loss. Accurate and early prediction of this disease is crucial for effective prevention and mitigation strategies. Traditional prediction methods often struggle with the complex interplay of environmental factors, host resistance, and pathogen dynamics, limiting their practical application. This paper aims to bridge this gap by developing a data-driven, AI-based system that can forecast rice blast incidence a year ahead, considering regional variations and specific rice cultivars.

Methods

The core of this research lies in the application of Long Short-Term Memory (LSTM) recurrent neural networks (RNNs). LSTMs are a type of RNN specifically designed to capture long-term dependencies in sequential data, making them well-suited for time-series forecasting. The authors developed region-specific LSTM models for three major rice-producing areas in South Korea: Cheolwon, Icheon, and Milyang.

The prediction process involves training the LSTM models using historical data from the preceding three years to predict the disease incidence for the fourth year. The model's performance was evaluated by systematically varying the input variables. These variations included:

  • Blast_LSTM: Using only historical rice blast disease scores.
  • BlastT_LSTM: Incorporating blast scores and average air temperature.
  • BlastTH_LSTM: Adding relative humidity to the previous inputs.
  • BlastTHS_LSTM: Including blast scores, temperature, humidity, and sunshine hours.
  • Climate_LSTM: Using only climatic data (temperature, humidity, sunshine hours).

The disease scores were classified into three categories: resistance (0-3), moderate resistance (4-6), and susceptibility (7-9). Data normalization (min-max scaling) was applied to all input variables to ensure consistent processing.

Datasets

The study utilized two primary types of datasets:

  • Rice Blast Disease Score Data: Collected by the National Institute of Crop Science, RDA, in South Korea from 2003 to 2016. This data comprises field trial results for 358 rice cultivars across 12 regions, with disease incidence graded on a 0-9 scale. For this study, data from four major rice-growing regions (Cheolwon, Icheon, Milyang, and Naju) were selected.
  • Historical Climatic Data: Daily data for average temperature, relative humidity, and sunshine hours for June and July from 2003 to 2016, obtained from the Korean Meteorological Agency. This data was aggregated into 15-day periods to match the time steps of the blast score data.

The study also analyzed the prediction performance for 17 of the most popular rice cultivars in Korea, using a test dataset derived from the broader cultivar data.

Key Insights and Ideas

  • LSTM's Efficacy for Temporal Patterns: The study strongly supports the use of LSTMs for capturing the temporal dynamics of rice blast disease, outperforming simpler models that don't account for sequential dependencies.
  • Synergistic Impact of Input Variables: Combining historical blast scores with climatic data significantly improves prediction accuracy. The most comprehensive model (BlastTHS_LSTM), incorporating all four input variables, consistently yielded the highest accuracies across regions and cultivars.
  • Regional Specificity Matters: Developing region-specific models is crucial, as climatic conditions and their impact on disease incidence vary significantly across different geographical areas.
  • Cultivar-Agnostic Prediction as a Foundation: While the ultimate goal is cultivar-specific prediction, the study demonstrates that a cultivar-agnostic model using historical blast scores and climate data can provide a strong baseline for predicting future outbreaks.
  • Early Warning System Potential: The ability to predict disease incidence a year in advance offers a substantial advantage for proactive management, allowing farmers and researchers to prepare and implement preventative measures.

Key Results

  • Improved Accuracy with Combined Inputs: The BlastTHS_LSTM model, using all four input variables (blast scores, temperature, humidity, sunshine hours), achieved the highest prediction accuracies. For instance, in Cheolwon, it reached 67.4% accuracy and a 63.6% F1-score.
  • Regional Performance Differences: Prediction accuracy varied by region, with Cheolwon generally showing the highest performance, followed by Icheon, and then Milyang. This highlights the importance of regional modeling.
  • Cultivar-Specific Insights: When evaluating the top 17 popular cultivars, the BlastTHS_LSTM model achieved an average accuracy of 79.4% for Cheolwon, 64.7% for Icheon, and 55.6% for Milyang. While not perfect, these results indicate the potential for cultivar-specific forecasting.
  • Climate Data's Contribution: Even when used alone (Climate_LSTM), climate data provided some predictive capability, though significantly lower than models incorporating blast scores. This confirms climate as a critical factor.
  • Baseline Performance: The Blast_LSTM model, using only historical blast scores, achieved respectable accuracies (e.g., 62.3% in Cheolwon), demonstrating that past disease incidence is a strong predictor.

Limitations and Future Work

  • Data Availability and Missing Values: The study encountered missing data, necessitating data removal and potentially limiting the scope of analysis. The authors note that using data from all regions would have significantly reduced usable data.
  • Model Accuracy: While improved, the prediction accuracies are not exceptionally high. The authors acknowledge this and suggest that performance can be further enhanced by:
  • Increasing Training Data: More extensive historical data would likely improve the LSTM's learning capabilities.
  • Optimizing LSTM Architecture: Fine-tuning hyperparameters and exploring more complex LSTM architectures could lead to better results.
  • Incorporating More Variables: Including other relevant factors like soil conditions, fertilizer application, and specific pathogen strains could further refine predictions.
  • Variety-Specific Data Size: The analysis for individual cultivars had a small sample size (1-4 instances), which might affect the reliability of those specific predictions.
  • Transfer Learning Potential: The authors highlight that the deep learning approach used is amenable to transfer learning, suggesting that the models could be adapted for use in other countries or regions with similar rice cultivation practices. This opens avenues for global application.

Summary updated

Abstract

Among all diseases affecting rice production, rice blast disease has the greatest impact. Thus, monitoring and precise prediction of the occurrence of this disease are important; early prediction of the disease would be especially helpful for prevention. Here, we propose an artificial-intelligence-based model for rice blast disease prediction. Historical data on rice blast occurrence in representative areas of rice production in South Korea and historical climatic data are used to develop a region-specific model for three different regions: Cheolwon, Icheon and Milyang. A rice blast incidence is then predicted a year in advance using long-term memory networks (LSTMs). The predictive performance of the proposed LSTM model is evaluated by varying the input variables (i.e., rice blast disease scores, air temperature, relative humidity and sunshine hours). The most widely cultivated rice varieties are also selected and the prediction results for those varieties are analyzed. Application of the LSTM model to the accumulated rice-blast disease score data confirms successful prediction of rice blast incidence. In all regions, the predictions are most accurate when all four input variables are combined. Rice blast fungus prediction using the proposed LSTM model is variety-based; therefore, this model will be more helpful for rice breeders and rice blast researchers than conventional rice blast prediction models.
BibTeX
@inproceedings{kim2017early,
  author = {Yangseon Kim and Jae-Hwan Roh and Ha Young Kim},
  title = {{Early Forecasting of Rice Blast Disease Using Long Short-Term Memory Recurrent Neural Networks}},
  booktitle = {Sustainability},
  year = {2017},
  doi = {10.3390/su10010034},
}
Figure