Iterative Training Sample Augmentation for Enhancing Land Cover Change Detection Performance With Deep Learning...

Area of Science:

• Remote Sensing
• Geospatial Analysis
• Artificial Intelligence

Background:

• Accurate land cover change detection (LCCD) using remote sensing images is crucial.
• Deep learning models require extensive labeled data for effective training.
• Manual sample labeling for bitemporal remote sensing images is laborious, time-consuming, and requires expertise.

Purpose of the Study:

• To propose an iterative training sample augmentation (ITSA) strategy to improve LCCD performance with deep learning.
• To reduce the reliance on manual sample labeling in LCCD tasks.
• To enhance the efficiency and accuracy of deep learning-based LCCD.

Main Methods:

• Developed an iterative training sample augmentation (ITSA) strategy.
• Measured sample similarity with neighboring blocks to identify potential new samples.
• Iteratively trained a neural network with augmented samples and predicted results.
• Coupled ITSA with existing deep learning networks for LCCD.

Main Results:

• The proposed ITSA strategy effectively improved LCCD performance when integrated with deep learning networks.
• Experiments demonstrated significant improvements in detection accuracy, with overall accuracy increases ranging from 0.38% to 7.53% compared to state-of-the-art methods.
• The ITSA strategy showed robust, generic, and universally adaptive performance across various image types and deep learning architectures.

Conclusions:

• The ITSA strategy is a valuable approach for enhancing deep learning-based LCCD.
• It effectively addresses the challenges of manual sample labeling, reducing effort and time.
• The method offers a significant and reliable improvement in LCCD accuracy, applicable to diverse remote sensing scenarios.