A Time Series Data Filling Method Based on LSTM-Taking the Stem Moisture as an Example.

Area of Science:

• Agricultural Science
• Data Science
• Machine Learning

Background:

• Sensor data collection in agriculture often suffers from data loss, impacting analysis and decision-making.
• Accurate imputation of missing time-series data is crucial for understanding plant physiological responses.

Purpose of the Study:

• To evaluate the accuracy and validity of different data imputation methods for plant stem moisture sensor data.
• To investigate the effectiveness of Long Short-Term Memory (LSTM) neural networks in filling missing time-series data.

Main Methods:

• Compared interpolation, statistical methods, Recurrent Neural Network (RNN), and LSTM for data imputation.
• Utilized stem moisture data from Lagerstroemia Indica, with manually deleted data segments.
• Developed a multi-dimensional LSTM model incorporating environmental parameters (humidity, radiation, soil temperature) for improved imputation.

Main Results:

• Bidirectional LSTM models demonstrated superior accuracy, achieving a mean absolute percent error (MAPE) of 1.813% for one-dimensional imputation.
• Multi-dimensional LSTM imputation, using environmental data, further enhanced accuracy with a minimum MAPE of 1.499%.
• Multi-dimensional filling successfully extended sequence length without accumulating errors, unlike one-dimensional methods.

Conclusions:

• LSTM-based data imputation offers significant advantages for filling long-lost time-series sensor data in plant science.
• The multi-dimensional approach effectively addresses error accumulation issues inherent in one-dimensional time-series imputation.
• This research provides a novel approach for robust data imputation in agricultural sensor networks.