研究目的
To propose an automatic extraction method for forest based on self-organization map (SOM) using GF-2 images to improve efficiency and accuracy in forest area extraction for national forest survey projects.
研究成果
The proposed automatic extraction method based on SOM achieves high classification accuracy (94.42% to 98.88%) for forest area extraction from GF-2 images, meeting the requirements of the national forest survey project. It reduces human error and does not require extensive prior knowledge, but future work should integrate supervised learning for better class matching and test in more complex areas.
研究不足
The method has difficulty separating forest from grassland due to similar input vectors, and building shadows in urban areas affect accuracy. It is strongly dependent on sample quality, and output classes may not always match expected categories without additional supervised learning.
1：Experimental Design and Method Selection:
The study uses a self-organizing map (SOM) algorithm, an unsupervised artificial neural network, for automatic forest extraction from high-resolution GF-2 satellite images. The method involves normalization of input vectors, search for winning neurons, and weight vector adjustment based on a 'winner-take-all' rule.
2：Sample Selection and Data Sources:
GF-2 satellite images acquired on September 6, 2017, were used, covering three study areas in Hebei province and Beijing city with varying forest cover (heavy, medium, low). The images underwent radiometric calibration, atmospheric correction, geometric correction, and pan-sharpening to 1-meter resolution.
3：List of Experimental Equipment and Materials:
GF-2 satellite images, MATLAB software with SOM module.
4：Experimental Procedures and Operational Workflow:
The process includes building an initial neural network, normalizing input and weight vectors, iteratively finding winning neurons, adjusting weights, and outputting classification results until convergence. The SOM was configured with hexagonal topology, Euclidean distance, input dimension of 4 (for multispectral bands), and output dimension of
5：Data Analysis Methods:
Classification accuracy was assessed using confusion matrices by comparing results with human-computer interactive interpretation, calculating accuracy percentages for each study area.
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
