Classification of satellite images for ecology management using deep features obtained from convolutional neural network models

• Published in: Iran Journal of Computer Science (Online) 2023-09, Vol.6 (3), p.185-193
• Main Authors: Özbay, Erdal, Yıldırım, Muhammed
• Format: Article
• Language: English
• Subjects: Artificial Intelligence; Computer Science; Computer Systems Organization and Communication Networks; Data Structures and Information Theory; Mathematics of Computing; Software Engineering/Programming and Operating Systems; Theory of Computation
• ISSN: 2520-8438; 2520-8446
• DOI: 10.1007/s42044-022-00133-6

Ecology is the scientific study of balancing biodiversity, which has an impact on natural life and habitats and establishes a strong yet complicated link between the ecosystem’s components. Climate change, wildlife, and other habitats are adversely affected by the presence of anthropogenic pressure. In this respect, it is very important to protect and map natural resources to create efficient ecology management models. In addition, the most efficient method for determining the natural resources and platforms on earth is satellite image analysis. It is effective in monitoring biological diversity, such as ecology management, environmental planning, forestry, agriculture, surface changes, and land use with satellite images. Current classification approaches using satellite imagery often have limited capabilities with feature coding producing mediocre results. Image classification has become quite effective with the development of deep learning models. This study aims to improve the classification performance of deep learning models in satellite image analysis for ecology management using image processing techniques. To manage the classification process more efficiently, convolutional neural network (CNN) models and the neighborhood component analysis (NCA) are used together. Unnecessary features are eliminated with the NCA method. Then, the feature map optimized by the NCA method was used for classification. MobileNetV2, DenseNet201, and ResNet50 were used as feature extractors and six different machine learning classifiers were used as classifiers. As a result, the success rate of classification of satellite images using derived feature vectors has been revealed as 96.46%. According to the experimental results, the use of a combination of feature selection approaches and convolutional neural network models helped to successful classify satellite images.