Classification of Satellite Images Using an Ensembling Approach Based on Deep Learning

The classification of satellite images is crucial for a wide range of applications. These applications require manual work to classify each image and label them correctly. Despite this, existing satellite image classification methods do not provide satisfactory results, and their performance is flawed. To address these requirements, we propose a deep learning-based approach to classify and label satellite images. To train our architectures, we used RSI-CB128, a crowd-sourced geographic dataset with 36,707 images distributed among 45 classes. We used different deep learning techniques like transfer learning and adding a few fully connected layers like GlobalAveragePooling, and Dense layers with activation functions such as Softmax and ReLU. Some architectures are selected for experiment, and the best four models among them are paired for the ensembling technique. Compared to similar state-of-the-art methods, our proposed best-ensembled model performs better and produces 99.30% classification accuracy and 0.992 F1-score, demonstrating its effectiveness in classifying satellite images.