Integrating Chaotic Optimization and Machine Learning Algorithms for Precise Remote Sensing Maps of Oakland, California, using Sentinel-2 Multispectral Imagery

The advancement of remote sensing technology has revolutionized urban planning and environmental monitoring, providing high-resolution spatial data essential for managing complex metropolitan areas like Oakland, California. However, the primary obstacle in achieving high-accuracy land cover classification is spectral heterogeneity. To address this, our research develops an advanced urban land cover mapping framework for an area of ​​51.88 km^2 in Oakland, California, using Sentinel-2 multispectral imagery. The primary objective is to enhance the performance of a Random Forest (RF) classifier by integrating chaotic optimization algorithms to fine-tune hyperparameters, thereby overcoming the limitations of traditional grid-search methods. Our study focuses on six distinct urban classes: High-Density Compact Urban, Large Low-Rise, Residential Open Low-Rise, Hillside Residential, Mixed-Use Transit Corridors, and Coastal Infrastructure. First, we developed and compared the hyperparameters and classification accuracy of a standard Random Forest Algorithm against a Chaotic Random Forest (CRF) variant. This comparison was conducted using three distinct built-up indices independently: the Modified Normalized Difference Built-up Index (MNDBI), the Built-up Area Extraction Index (BAEI), and the Normalized Difference Impervious Surface Index (NDISI). Second, the optimized parameters were applied to generate individual built-up classification maps for each index, assessing their sensitivity to Oakland’s unique topography. Finally, a proposed fusion approach synthesized these results with additional environmental layers such as the Normalized Difference Vegetation Index (NDVI), the Modified Normalized Difference Water Index (MNDWI), and the Barren Soil Index (BSI) to create a final high-precision land cover map. Preliminary results indicate that the Chaotic Random Forest (CRF) model significantly outperforms the standard Random Forest (RF) model. Specifically, the CRF achieved an Overall Accuracy (OA) of 80.2% and a Kappa coefficient of 0.92 compared to the standard RF’s OA of 76% and Kappa of 0.84.