Machine Learning Approaches for Developing Land Cover Mapping.

Area of Science:

• Remote Sensing
• Machine Learning
• Geospatial Analysis

Background:

• Urban land cover classification is crucial for environmental management and sustainable development.
• Existing methods often use orthographic imagery and digital surface models (DSMs).
• Feature extraction from high-resolution satellite images is key, but irrelevant features can reduce accuracy.

Purpose of the Study:

• To improve urban land cover classification accuracy in remote sensing.
• To investigate the effectiveness of a genetic algorithm-based feature selection approach.
• To evaluate the performance of neural networks (NNs) and random forest (RF) classifiers with selected features.

Main Methods:

• A genetic algorithm was employed for feature selection from high-resolution satellite data.
• Neural networks and random forest classifiers were used to assess the selected features.
• The approach was tested on a dataset comprising nine urban land cover classes.

Main Results:

• The genetic algorithm-based feature selection enhanced classification performance.
• The random forest classifier achieved the highest accuracy (84.27%) using only 27% of the original features.
• This demonstrates the effectiveness of reducing feature dimensionality for improved classification.

Conclusions:

• Feature selection using genetic algorithms is effective for urban land cover classification.
• The random forest algorithm is well-suited for this task, offering high accuracy with reduced feature sets.
• Optimizing feature selection is vital for efficient and accurate remote sensing data analysis.