研究目的
Mapping 2017 early rice using a CNN in which the parameters are transferred from a successfully pre-trained CNN and fine-tune it with slight adjustment.
研究成果
The experiment validated that with fine-tuning of a small amount of data, the CNN model using middle-resolution remote sensing data could be transferred to suit the early rice mapping task in different time periods, achieving an overall accuracy of 81.68%.
研究不足
The CNN algorithm is extremely data-hungry, requiring significant training data for each mapping task if not using transfer learning.
1：Experimental Design and Method Selection
The study proposes a transfer learning method that pre-trains a CNN with middle-resolution remote sensing data in 2016 and fine-tunes it in 2017 with high-resolution remote sensing data.
2：Sample Selection and Data Sources
The research site is located in 25 counties, Jiangxi Province, China. The data source is 16-meter resolution images taken by GF-1 satellite platform, containing Red, Green, Blue, and near-infrared (NIR) bands.
3：List of Experimental Equipment and Materials
GF-1 satellite platform images.
4：Experimental Procedures and Operational Workflow
Images were filtered for three critical periods, cropped into square patches, and labeled using SVM classification. The pre-trained model was fine-tuned with 2017 data.
5：Data Analysis Methods
A confusion matrix was constructed for accuracy assessment, calculating overall accuracy, Kappa coefficient, and F1-score.
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
