研究目的
To test the performance of a new crop classification method (PSP) for fused Sentinel-1 and Sentinel-2 images and to compare it with the Random Forest classifier, as well as to show the gain from data fusion for crop classification.
研究成果
The PSP classification method outperforms Random Forest for both Sentinel-1 only and fused datasets, demonstrating its robustness for agricultural classification. Data fusion with Sentinel-2 can improve classification for certain crops when images are available at crucial phenological stages, but its benefits are limited by low data availability due to clouds.
研究不足
Low availability of Sentinel-2 data due to cloud cover, with only three images available during the vegetation period, and some images acquired after crop harvest, limiting the effectiveness of data fusion for most crops.
1：Experimental Design and Method Selection:
The study uses the Phenological Sequence Pattern (PSP) method for classification, which incorporates crop phenology information, and compares it with the Random Forest classifier. The methodology involves processing and fusing Sentinel-1 radar and Sentinel-2 multispectral images over one growing season.
2：Sample Selection and Data Sources:
The study area is around Hanover, Germany, covering 2400 km2 with agricultural land. Ground truth data includes 257 fields with 12 crop types, collected from October 2015 to October
3：List of Experimental Equipment and Materials:
20 Sentinel-1 GRDH products, Sentinel-2 Level-1C and Level-2A products, SNAP software with S1-Toolbox and S2-Toolbox, Sen2Cor Tool, SRTM digital elevation model, and R programming language (version 3.3.1) with packages sp, rgdal, raster, randomForest, and maptools.
4：1) with packages sp, rgdal, raster, randomForest, and maptools.
Experimental Procedures and Operational Workflow:
4. Experimental Procedures and Operational Workflow: Images were preprocessed for geocoding, radiometric calibration, atmospheric correction, and coregistration. For PSP classification, images were selected based on phenological stages, and probabilities were calculated using Random Forest. Non-PSP Random Forest used all features without phenological consideration.
5：Data Analysis Methods:
Classification accuracies were evaluated using F1-score and overall accuracy (OA). Results were compared between PSP and Random Forest, and between fused (S1+S2) and non-fused (S1 only) datasets.
独家科研数据包，助您复现前沿成果，加速创新突破
获取完整内容
