ENHANCED LAND USE AND LAND COVER CLASSIFICATION USING WEIGHTED ENSEMBLE LEARNING OF PRETRAINED CNN MODELS

Abstract

Accurate Land Use and Land Cover (LULC) classification is essential for environmental monitoring, precision agriculture, urban planning, and disaster management. Although Convolutional Neural Networks (CNNs) have significantly advanced remote sensing image interpretation, their individual performance varies across heterogeneous land-cover categories due to architectural differences and representational biases. To address these inconsistencies, this study proposes an enhanced ensemble learning framework that integrates six pretrained CNN architectures—VGG16, ResNet50, InceptionV3, DenseNet121, EfficientNetB0, and MobileNetV2—fine-tuned on the UC Merced Land Use dataset. Three ensemble strategies were examined: simple averaging, soft voting, and weighted averaging based on grid-search optimization. Extensive experiments were conducted to evaluate classification accuracy, F1-score, inference time, and GPU memory consumption. DenseNet121 achieved the highest individual accuracy (97.05%), while the weighted averaging ensemble significantly outperformed all standalone models, reaching 99.43% accuracy and an F1-score of 0.993 with efficient inference performance. These findings demonstrate that combining diverse CNN architectures enhances classification robustness, reduces model-specific weaknesses, and improves stability across visually similar LULC categories. The proposed ensemble framework offers a scalable, high-accuracy solution suitable for real-world Earth observation tasks and contributes a comprehensive comparative evaluation of ensemble strategies for LULC classification using deep transfer learning.