研究目的
To analyze the value of different modalities (spectral and shape information) for semantic segmentation of aerial imagery using a Residual Shuffling Convolutional Neural Network (RSCNN).
研究成果
The results indicate that true orthophotos are better suited for classification than DSM or geometric features alone. Combining radiometric and geometric information improves classification accuracy, with the best results achieved using IR, R, G, nDSM, NDVI, L, P, and S layers. The additional consideration of geometric features (linearity, planarity, sphericity) yields improvement, while vegetation indices like NDVI do not enhance results.
研究不足
The study is limited to the Vaihingen Dataset, which may not generalize to other regions or data types. The training and test data split differs from the standard ISPRS benchmark, potentially affecting comparability. Computational constraints from using a single GPU and specific patch sizes may limit scalability.
1：Experimental Design and Method Selection:
The methodology involves using multi-modal data (true orthophoto, DSM, and derived features) as input to a Residual Shuffling Convolutional Neural Network (RSCNN) for semantic segmentation. The RSCNN combines a ResNet with atrous convolution and a shuffling operator for dense prediction.
2：Sample Selection and Data Sources:
The Vaihingen Dataset is used, consisting of 33 patches with true orthophoto (IR, R, G bands), DSM, and reference labeling for six semantic classes. Training uses patches 1,3,5,7,13,17,21,23,26,32,37; evaluation uses patches 11,15,28,30,
3：List of Experimental Equipment and Materials:
NVIDIA TITAN X GPU with 12GB RAM, MXNet library for implementation.
4：Experimental Procedures and Operational Workflow:
Extract features (nDSM, NDVI, linearity L, planarity P, sphericity S) from the input layers. Train the RSCNN using stochastic gradient descent with momentum, with patches of varying sizes (56x56 to 448x448 pixels) for different epochs. Perform classification on different subsets of the nine layers.
5：Data Analysis Methods:
Evaluate performance using Overall Accuracy (OA), mean F1-score (mF1), mean Intersection-over-Union (mIoU), and class-wise F1-scores.
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
