A Discriminative Long Short Term Memory Network with Metric Learning Applied to Multispectral Time Series...

Area of Science:

• Computer Science
• Machine Learning
• Remote Sensing

Background:

• Multi-spectral time series analysis is crucial for crop type mapping.
• Long Short-Term Memory networks (LSTMs) are effective for time series data but struggle with high intra-class variance and inter-class similarity.
• Existing methods face challenges in distinguishing similar crop types and handling variations within a single crop type over time.

Purpose of the Study:

• To develop an end-to-end deep network for improved multi-spectral time series classification.
• To enhance crop type mapping accuracy by addressing challenges in temporal data analysis.
• To leverage metric learning to create more discriminative features for classification.

Main Methods:

• Proposed a novel deep network architecture integrating Long Short-Term Memory (LSTM) modules.
• Employed a metric learning approach with a triplet loss function to enforce feature discriminability.
• Utilized three distinct, shared-weight branches, each containing an LSTM, merged via the triplet loss.

Main Results:

• The proposed network effectively minimizes classification error in multi-spectral time series.
• The architecture successfully generates more discriminative deep features, improving classification performance.
• Achieved validated results on the challenging BreizhCrops dataset for crop type mapping.

Conclusions:

• The combination of LSTMs and metric learning offers a powerful approach for crop type mapping.
• The developed deep network architecture enhances the ability to classify complex multi-spectral time series data.
• This method shows significant promise for accurate and robust agricultural monitoring using remote sensing data.