Enhanced Land Cover Classification by Integrating Multispectral Satellite Imagery and Machine Learning Techniques

Abstract

For resource management, urban planning, and environmental monitoring, accurate land use and land cover (LULC) categorization is essential. Nevertheless, spectral similarity across land cover categories, feature redundancy, and the shortcomings of conventional classification methods make it difficult to achieve high classification accuracy. The majority of research uses either object-based (OB) or pixel-based (PB) classification, not making use of their complementing advantages. Furthermore, feature selection is frequently done by manual analysis which does not necessarily provide accurate classification results. This paper suggests a hybrid OB-PB classification strategy in which OB classification is used for fine-tuning mapping after PB classification is used for initial segmentation. Recursive feature elimination (RFE) is an automated feature selection method that replaces traditional manual selection of textural (GLCM) and statistical (PCA) features in order to further improve classification accuracy. The study's emphasis is Rajasthan, India, which spans 219,676.75 km² and is divided into five LULC types: desert, agricultural, water, forest, and urban. Four machine learning classifiers Random Forest (RF), Support Vector Machine (SVM), Classification and Regression Trees (CART), and Gradient Tree Boosting (GTB) were used with 249 training examples. The kappa coefficient and the confusion matrix were used to evaluate accuracy. The results show that the hybrid OB-PB technique performs better in terms of segmentation and classification precision than the independent OB and PB approaches. The classifier's eligibility for LULC mapping was confirmed by GTB's greatest accuracy among the others. For LULC classification, this paper offers a unique framework that combines automated feature selection with hybrid classification, providing a more precise, effective, and scalable method. Improved classification techniques and better decision-making for ecological surveillance and management of land are two benefits of the research.