Comparing Three Methods of Selecting Training Samples in Supervised Classification of Multispectral Remote Sensing...

Area of Science:

• Earth Observation
• Geospatial Analysis
• Image Classification

Background:

• Effective training sample selection is critical for accurate remote sensing image classification.
• Various methods exist for selecting training samples, each with potential impacts on classification performance.

Purpose of the Study:

• To compare the efficacy of three training sample selection methods: grouping selection, entropy-based selection, and direct selection.
• To evaluate the performance of random forest (RF), support-vector machine (SVM), and k-nearest neighbor (KNN) classification models using these selection methods on Sentinel-2, GF-1, and Landsat 8 imagery.

Main Methods:

• Utilized Sentinel-2, GF-1, and Landsat 8 satellite imagery.
• Implemented grouping selection, entropy-based selection, and direct selection for training data.
• Trained and evaluated RF, SVM, and KNN supervised classification models.

Main Results:

• All three classification models (RF, SVM, KNN) demonstrated similar performance across the evaluated imagery.
• The grouping selection method yielded higher classification accuracy than entropy-based selection, using fewer samples.
• Grouping selection also outperformed direct selection when an equivalent number of samples were used.

Conclusions:

• The grouping selection method is superior for training sample selection in remote sensing image classification.
• Optimal classification accuracy is achieved with the grouping selection method, especially within a specific range of sample sizes.