Land use and land cover (LULC) refer to how humans utilize land for various purposes, such as agriculture, forestry, and urban development. Land use encompasses the physical characteristics of the land, including vegetation, built-up areas, and water bodies. Monitoring LULC is essential for tracking changes in land patterns due to natural or human activities. It also aids in assessing risks such as deforestation, flooding, and landslides. Therefore, generating LULC maps is crucial for continuously monitoring these changes.
Traditional surveying methods for mapping land use can be expensive and labor-intensive. In recent decades, novel methods have emerged as valuable alternatives. One effective method for updating land use maps in both urban and non-urban areas is the processing of remote sensing data. Many past studies have introduced several techniques for change detection, including Artificial Neural Networks (ANN), Chi-Square transformation, image ratioing, image differencing, and hybrid change detection approaches.
Change detection techniques in remote sensing can be classified into five categories: algebra-based methods, transform-based methods, classification-based methods, GIS-based methods, and advanced methods. Key algebra-based models include image differencing, image ratioing, and Change Vector Analysis (CVA). Principal Component Analysis (PCA) is recognized as a prominent transform-based method. Notably, past studies have indicated that ANN, particularly feedforward networks, stands out as one of the most effective machine learning models for this purpose.
A study conducted by Mahdi Sedighkia and Bithin Datta from the College of Science and Engineering at James Cook University, Townsville, Australia, has explored methodologies for detecting land use and land cover (LULC) changes in Cairns, Queensland. The research employs both metaheuristic optimization techniques, namely Particle Swarm Optimization (PSO) and Invasive Weed Optimization (IWO), as well as conventional methods, including Levenberg-Marquardt Backpropagation (LM), Scaled Conjugate Gradient Backpropagation (SCG), and BFGS Quasi-Newton Backpropagation.
Focusing on the northern region of Queensland, the study highlights the significance of effective monitoring of agricultural lands—critical to the local economy, particularly in the context of sugar cane cultivation. The diverse ecosystems within this area encompass a variety of landscapes, including mountains, beaches, wetlands, and rivers. The findings underscore the necessity for reliable and cost-effective approaches to assess changes in land use and land cover, which are essential for effective
environmental management and planning in agricultural sectors.
The study employed an artificial neural network (ANN) architecture comprising eight hidden layers to develop machine learning models aimed at land use classification. The primary objective is to assess the effectiveness of metaheuristic optimization techniques in training these models. The focus is on two algorithms: Particle Swarm Optimization (PSO), a well-established method, and Invasive Weed Optimization (IWO), both of which are utilized to optimize the weights of the neural network. The findings indicate that PSO effectively demonstrates the capabilities of traditional algorithms in training classifiers for the identification of agricultural lands, while IWO illustrates the potential of newer generation algorithms to enhance the overall performance of machine learning models.
The authors of the study examine various models to evaluate their efficacy in detecting land use changes,
specifically within agricultural contexts. We employed key performance metrics, including the Overall Accuracy Index (OAI), User’s Accuracy (UA), Producer’s Accuracy (PA), and the Kappa coefficient, to provide a comprehensive assessment of accuracy. The research specifically investigates the performance of two metaheuristic optimization techniques—Particle Swarm Optimization (PSO) and Improved Wolf Optimization (IWO)—in conjunction with remote sensing data and a machine learning model utilizing the Feedforward Neural Network (FNN) classifier. These metaheuristic approaches were compared against three conventional training methods: Scaled Conjugate Gradient (SCG), Levenberg-Marquardt (LM), and BFGS Quasi-Newton Backpropagation (BFG).
The findings indicate that the conventional methods, particularly BFG, demonstrate superior performance in terms of both accuracy and computational efficiency compared to the metaheuristic techniques. This suggests that established training methodologies may be more effective for the development of robust machine learning models aimed at land use detection.
