High spatial resolution crop type and land use land cover classification without labels: A framework using multi-temporal PlanetScope images and variational Bayesian Gaussian mixture model
Previous studies often combined high spatial resolution data (e.g., PlanetScope) with wider spectral range data (e.g., Sentinel-2) and relied on supervised classification methods to produce land use and land cover (LULC) maps. This study proposed a new unsupervised framework to generate crop type an...
Saved in:
| Main Authors: | , , , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
Elsevier
2025-12-01
|
| Series: | Science of Remote Sensing |
| Subjects: | |
| Online Access: | http://www.sciencedirect.com/science/article/pii/S2666017225000707 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Previous studies often combined high spatial resolution data (e.g., PlanetScope) with wider spectral range data (e.g., Sentinel-2) and relied on supervised classification methods to produce land use and land cover (LULC) maps. This study proposed a new unsupervised framework to generate crop type and LULC maps at high spatial resolution (<5 m) using available PlanetScope data solely without requiring ground truths. We used PlanetScope surface reflectance images and their derived spectral indices during growing seasons to create multi-temporal input features, which were fed into an unsupervised Variational Bayesian Gaussian Mixture Model (VBGMM). The VBGMM, unlike the traditional unsupervised classification methods, (1) first estimated optimal parameters that are most suitable based on the input features and then (2) assigned pixels to the cluster with maximum posteriori probability of a mixture of several Gaussian distributions. The crop type and LULC maps were then generated by labeling the derived clusters using the best possible assignment method, referring to the existing crop type or LULC products. We evaluated the produced PlanetScope-based crop type and LULC maps using true labels, corresponding reference maps, and other unsupervised classification methods. The results demonstrated the robustness and effectiveness of the proposed framework in mapping crop types and LULC at 3–5 m pixels across various ecosystems, climate zones, and human-managed landscapes. The spatial patterns of PlanetScope-based maps were (1) highly comparable with all the reference datasets at 10–30 m spatial resolution and (2) better than the traditional GMM and K-means clustering methods. The VBGMM produced classification maps with high confidence, yielding class probabilities above 0.9 for over 90 % of all study areas. The area percentage for all crop type and LULC classes agreed well with their reference maps, with R2 of 0.95 and RMSE of 1.04 %. The confusion matrices using true labels indicated that PlanetScope-based maps achieved a higher overall accuracy of 84 % than the supervised referenced maps of 81 %. Besides, the entropy comparison showed that our framework-based maps were better at capturing fine-scale features such as developed areas within cities that commonly mix with open space and vegetation, deforestation and cropland conversion in South America, smallholder croplands in Africa and Asia, and generating homogeneous crop fields in North America. This study further highlighted the potential for future research to implement our proposed framework to generate timely and extensive annotated datasets, which can be used for operationally training machine learning models to map crop types and LULC, track deforestation, detect wildfires, and delineate flooded areas at larger scales using medium/coarse Earth observations. |
|---|---|
| ISSN: | 2666-0172 |