Time Series Remote Sensing Image Classification with a Data-Driven Active Deep Learning Approach.

Area of Science:

• Earth and Environmental Sciences
• Computer Science
• Artificial Intelligence

Background:

• Time Series Remote Sensing Images (TSRSIs) are crucial for land use/land cover (LULC) mapping.
• Deep learning excels at processing temporal data but requires extensive labeled samples.
• Manual labeling of TSRSIs is time-consuming and labor-intensive.

Purpose of the Study:

• To develop an active deep learning framework for TSRSI classification.
• To address the challenge of limited labeled data in TSRSI analysis.
• To reduce human effort in labeling large-scale remote sensing datasets.

Main Methods:

• Proposed a data-driven active deep learning framework for TSRSI classification.
• Designed a temporal classifier and an active learning strategy considering representativeness (K-shape clustering) and uncertainty (auxiliary deep network).
• Introduced a novel loss function to enhance deep model performance.

Main Results:

• Achieved significant accuracy improvements on multiple TSRSI datasets (MUDS, DynamicEarthNet, PASTIS).
• Demonstrated substantial gains, e.g., 7.81% accuracy improvement on DynamicEarthNet with limited initial samples.
• The proposed active learning method effectively identifies informative samples, outperforming other approaches.

Conclusions:

• The developed active deep learning framework offers an efficient solution for TSRSI classification with limited labeled data.
• The method effectively balances sample representativeness and uncertainty for optimal active learning.
• This approach significantly reduces the cost and effort associated with large-scale remote sensing data labeling.