Modified Gaussian Mixture Model and EfficientNetV2 for Enhanced Remote Sensing Recognition

In recent scenarios, remote sensing image classification has gained significant attention for applications such as land use analysis, urban planning and environmental monitoring. However, traditional approaches often face limitations in capturing complex spatial patterns and struggle with computational efficiency. To address these challenges, this research proposes a Modified Gaussian Mixture Model (GMM) with an EfficientNetV2 framework for accurate aerial image classification. Initially, data is collected from the UCM dataset, which consists of aerial images categorized into 21 land-use classes, each containing 100 images. Preprocessing is performed using contrast stretching and homomorphic filtering to enhance features and normalize illumination. Next, a Graph-Cut optimized Fuzzy GMM (Modified GMM) is used to segment images precisely by integrating probabilistic clustering with edge-aware optimization. Diverse features are then extracted using azimuthal averaging, Haar wavelet transforms, and MSER, capturing textural, spectral, and shape-based information. These features are further optimized using Particle Swarm Optimization (PSO) to reduce dimensionality. Finally, EfficientNetV2 is employed as the classifier due to its balanced depth, width, and resolution scaling, which enables high accuracy with low computational cost. The proposed Modified GMM with EfficientNetV2 framework achieves superior results in terms of accuracy of 96.71% compared to existing methods such as Modified GMM Segmentation and DenseNet.