Enhancing satellite image compositing with temporal proximity weighting for deep learning–based cropland segmentation
Generating composite images from satellite data is crucial for crop mapping over defined periods. However, producing reliable composites for cropland segmentation presents challenges, particularly in maintaining temporal coherence and preserving key phenological stages in time series data. This stud...
Saved in:
| Main Authors: | , , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
Elsevier
2025-09-01
|
| Series: | International Journal of Applied Earth Observations and Geoinformation |
| Subjects: | |
| Online Access: | http://www.sciencedirect.com/science/article/pii/S1569843225004510 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Generating composite images from satellite data is crucial for crop mapping over defined periods. However, producing reliable composites for cropland segmentation presents challenges, particularly in maintaining temporal coherence and preserving key phenological stages in time series data. This study proposes a compositing method that improves temporal coherence for tracking phenological stages in deep learning–based cropland segmentation. The compositing method integrates the near–infrared to blue band reflectance ratio with a Gaussian weighting function to prioritize pixel selection based on temporal proximity to the center of the target month. Sentinel–2 monthly time series composites were generated using Google Earth Engine and evaluated through proximity analysis to assess pixel distribution within the target month and correlations with consecutive months. The performance of deep learning models trained on these composites was further assessed by comparing their segmentation results. To evaluate generalizability, the method was applied across various study areas and across different crop types and environmental conditions. The results consistently show that proposed method outperforms other techniques in preserving temporal continuity, reducing cloud–related noise, and maintaining the coherence necessary for deep learning models to effectively track crop growth patterns. |
|---|---|
| ISSN: | 1569-8432 |