PSO-Optimized CNN for Feature Extraction and Accurate Classification of Satellite Images Using Machine Learning

Satellite imagery is indeed a powerful tool for various applications, including environmental monitoring, infrastructure planning and management, urban planning, and disaster response. This technology allows for a comprehensive overview of the Earth's surface, facilitating the remote identification and examination of transformations that may be challenging to observe from a ground-level perspective. However, traditional object classification and detection algorithms can sometimes be unreliable and inaccurate when processing satellite imagery. This can be due to a variety of reasons, such as limited resolution and noise in the imagery, variations in lighting and atmospheric situations, and the complication of the objects and its features being detected. Over past few years, techniques that rely on deep learning have demonstrated impressive achievements in enhancing the dependability and precision of object classification and detection in satellite imagery intended for land classification purposes. These techniques use PSO to optimize hyperparameters for convolutional neural networks (CNNs) to extract features from the images and identify objects based on learned patterns., and SVMs can be used to classify the extracted features. These proposed techniques can be combined to create powerful and accurate image classification models with an accuracy of 95%.

I. Introduction

Many years ago, land use and land cover classification has been an essential field of study within remote sensing and geospatial analysis. The data related to land use and land cover is highly valuable for understanding structure and changes taking place on the Earth's surface. This information has a diverse range of applications, including but not limited to urban planning, environmental management, disaster assessment, and agricultural monitoring. In the past, remote sensing imagery based (LULC) land use and land cover classification methods depended on manually designed characteristics and decision trees. However, these methods lacked accuracy and were time-consuming to implement. Over the past few years, deep learning algorithms techniques have surfaced as a promising solution for this land use and land cover classification problems.