MPGH-FS: A Hybrid Feature Selection Framework for Robust Multi-Temporal OBIA Classification | JoVE Visualize

Area of Science:

Remote Sensing
Image Analysis
Computer Science

Background:

Object-Based Image Analysis (OBIA) in high-resolution remote sensing often suffers from the curse of dimensionality due to high-dimensional features.
This leads to reduced classification efficiency and generalizability.
Existing single-criterion feature selection methods have limitations in temporal adaptability.

Purpose of the Study:

To propose a novel feature selection framework, MPGH-FS, addressing the curse of dimensionality and enhancing temporal adaptability in remote sensing image classification.
To integrate Mutual Information Correlation Coefficient (MICC) pre-filtering, Genetic Algorithm (GA) global search, and Hill Climbing (HC) local optimization for robust feature selection.
To evaluate the framework's performance in terms of feature compression, classification accuracy, and cross-temporal transferability.

Main Methods:

Developed the Mutual information Pre-filtering and Genetic-Hill climbing hybrid Feature Selection (MPGH-FS) framework.
Integrated MICC pre-filtering for initial feature screening.
Employed GA for global search and HC for local optimization of feature subsets.
Conducted experiments using multi-temporal GF-2 satellite imagery (2018-2023).

Main Results:

MPGH-FS reduced feature dimensions from 232 to 9.
Achieved high Overall Accuracy (OA) of 85.55% and Kappa coefficient of 0.75 in full-scene classification.
Demonstrated efficient training (6 s) and inference (1 min) times.
Exhibited robustness to inter-annual variations, with classification accuracy fluctuations below 4% in cross-temporal transfer experiments.

Conclusions:

MPGH-FS effectively compresses features, enhances classification performance, and improves temporal adaptability in multi-temporal remote sensing.
The proposed framework offers a robust solution for efficient and accurate remote sensing image classification.
MPGH-FS overcomes limitations of single-criterion selection and enhances generalizability across different time periods.