Spatially Explicit Active Learning for Crop-Type Mapping from Satellite Image Time Series.

Area of Science:

• Remote Sensing
• Machine Learning
• Agricultural Science

Background:

• Supervised crop classification from remote sensing images faces challenges due to the high cost and time required for sample annotation.
• Traditional Active Learning (AL) methods often neglect the spatial context present in remote sensing data.
• Efficient sample selection is crucial for optimizing the performance of supervised classification models.

Purpose of the Study:

• To develop and evaluate a novel spatially explicit Active Learning (AL) method for crop type classification.
• To leverage semi-variogram analysis to identify and eliminate redundant, spatially proximate samples.
• To assess the efficiency and accuracy of the proposed AL method compared to traditional AL approaches.

Main Methods:

• Implementation of a spatially explicit AL strategy incorporating semi-variogram analysis to discard redundant samples.
• Utilized Random Forest (RF) classifier with Sentinel-2 Satellite Image Time Series data.
• Evaluated the method in two distinct study areas in the Netherlands and Belgium.

Main Results:

• In the Netherlands, spatially explicit AL required fewer samples (97) with comparable accuracy (80%) to traditional AL (169 samples, 82% accuracy).
• In Belgium, spatially explicit AL used fewer samples (223) achieving 60% accuracy versus traditional AL (327 samples, 63% accuracy).
• The method demonstrated effectiveness for distinct crop classes like sugar beets and cereals but faced challenges with aggregated classes.

Conclusions:

• The developed spatially explicit AL method offers an efficient approach to sample selection for crop classification.
• This method reduces annotation effort while achieving competitive classification accuracy.
• Further research may be needed to address challenges with classifying aggregated crop types.