研究目的
To accurately extract cropland parcels from very high-resolution remotely sensed imagery for precision agriculture and other fields.
研究成果
The proposed deep extraction method accurately maps smallholder cropped area over a large region by extracting hard edges and soft edges respectively and filtering out the cropland parcels with semantic segmentation. This deep learning based framework showed great potential to balance the large-scale and fragmental demands in agricultural remote sensing.
研究不足
The current semantic models have good generalized performance, but to remote sensing images with regional heterogeneity that is not enough. Larger area and more detailed experiments may help to understand and balance this contradiction better.
1：Experimental Design and Method Selection:
The study proposes a deep-edge guided method for cropland parcels extraction, focusing on the boundaries of these parcels. Hard edge and soft edge are extracted respectively with U-Net and RCF model.
2：Sample Selection and Data Sources:
The method was tested in Zongyang, Anhui Province of East China using images acquired from GF-2 satellite with a pansharpened resolution of
3：8m. List of Experimental Equipment and Materials:
GF-2 satellite images, U-Net and RCF models.
4：Experimental Procedures and Operational Workflow:
The method is divided into three major stages: sample preparing and training, edge and cropland predicting, and post-processing for complete cropland parcels.
5：Data Analysis Methods:
The study performs IoU-recall curve to address how much detected parcels could be used in GIS map.
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
