研究目的
To develop a machine learning framework that can learn from all available remote sensing scene datasets for improved classification accuracy, formulated as a multitask learning problem.
研究成果
The proposed multitask deep convolutional neural network framework shows promising capabilities in improving classification accuracy by learning from multiple remote sensing scene datasets. Preliminary results indicate that multitask classification provides improved overall accuracy compared to separate-task classification. However, further research is needed to optimize the framework.
研究不足
The study acknowledges the need for further research to improve classification, such as designing other architectures for better information sharing between tasks and reducing the discrepancy between distributions of different tasks.
1：Experimental Design and Method Selection:
The study employs a multitask learning framework using deep convolutional neural networks (CNNs), specifically SqueezeNet, for scene classification in remote sensing images. The framework is designed to share information between tasks (datasets) to improve classification accuracy.
2：Sample Selection and Data Sources:
Three datasets are used: UC Merced, AID, and KSA datasets, each representing different tasks with varying classes and conditions.
3：List of Experimental Equipment and Materials:
The experiments are conducted on an HP-station with an Intel Xeon processor 2.40GHz, 24.00 GB of RAM, and the GPU GEForce GTX1090 with 11GB of memory. The Keras environment is used for implementing the multitask network.
4：40GHz, 00 GB of RAM, and the GPU GEForce GTX1090 with 11GB of memory. The Keras environment is used for implementing the multitask network.
Experimental Procedures and Operational Workflow:
4. Experimental Procedures and Operational Workflow: The network is trained with data from all tasks jointly, with options to fix or make trainable parts of the pre-trained CNN (SqueezeNet). Training parameters include 100 epochs, a mini-batch size of 100 samples, and the Adam optimization method with a learning rate of 0.
5：Data Analysis Methods:
0001.
5. Data Analysis Methods: Performance is evaluated using overall accuracy (OA), calculated as the ratio of correctly classified samples to the total number of tested samples. Experiments are repeated 5 times with randomly selected training sets to compute average OA.
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
