研究目的
To overcome the limitations of conventional pixel-based contextual information extraction methods in VHR RS image classification by introducing novel spatial feature extractors (SPMPs and SPMPsM) and evaluating the new decision forest algorithm ForestPA.
研究成果
The proposed SPMPs and SPMPsM are effective for VHR RS image classification, with SPMPsM performing better in certain scenarios. ForestPA outperforms only bagging but is computationally inefficient and less accurate than other ensemble methods like ExtraTrees and Random Forest for high-dimensional data. Future work includes adaptive parameter selection and high-performance computing for acceleration.
研究不足
ForestPA has slow training efficiency, especially with high-dimensional data, and may not perform well with large sample sizes or high inter-band correlation. The number of superpixels and scale step size require tuning, which depends on image and object characteristics.
1：Experimental Design and Method Selection:
The study employs a comparative investigation using three VHR multi-/hyperspectral RS image datasets. Methods include morphological profiles (MPs), MPs with partial reconstruction (MPPR), and the proposed superpixel-guided morphological profiles (SPMPs and SPMPsM). Classifiers used are SVM, bagging, AdaBoost, MultiBoost, ExtraTrees, Random Forest, Rotation Forest, and ForestPA.
2：Sample Selection and Data Sources:
Three datasets are used: ROSIS Pavia University hyperspectral image (610x340 pixels, 103 bands), GRSS-DFC2013 hyperspectral image (340x1350 pixels, 144 bands), and Zurich QuickBird multispectral image (1295x1364 pixels, 4 bands). Training and test samples are predefined for each dataset.
3：List of Experimental Equipment and Materials:
Remote sensing images from ROSIS sensor, NSF-funded Center for Airborne Laser Mapping, and QuickBird satellite. Software for image processing and classification, including algorithms for superpixel generation (SLIC) and classifiers.
4：Experimental Procedures and Operational Workflow:
Generate MPs and MPPR using disk-shaped SEs with sizes from 1 to
5：Generate superpixels using SLIC algorithm with varying numbers and scale steps. Extract features (MPs, MPPR, SPMPs, SPMPsM) and apply PCA for dimensionality reduction. Train and test classifiers with normalized data, evaluate using overall accuracy, kappa statistic, and CPU time. Data Analysis Methods:
Statistical evaluation of classification accuracy (OA, kappa), computational efficiency (CPU time), and visual interpretation of classification maps. Parameters are tuned via cross-validation for SVM.
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
