Combining machine learning and adaptive large neighbourhood search for optimizing urban land use to reduce urban land surface temperature

Abstract Urban computing powered by high-resolution spatial data offers promising solutions for mitigating complex thermal issues in rapidly developing cities. This study presents a hybrid optimization framework that integrates the generative adversarial network (GAN) with an adaptive large neighbourhood search (ALNS) algorithm to optimize urban land use layouts for land surface temperature (LST) reduction. The GAN model is trained to learn spatial correlations between urban land use configurations and thermal distributions, enabling fast and flexible LST prediction from layout inputs. Building on this predictive capability, the ALNS algorithm explores a modularized nine-dimensional parameter space representing key urban land elements – such as vegetation, buildings, roads, and water bodies – to iteratively search for optimized configurations. Unlike traditional single-variable or standalone metaheuristic approaches, this framework captures complex non-linear spatial interactions and escapes local optima through adaptive operator selection. The optimization outcomes are validated through comparative experiments, demonstrating superior performance over baseline methods in identifying land use layouts with improved thermal characteristics. Additionally, the framework supports interactive visual mapping between inputs and outputs, allowing integration into early-stage urban planning processes. This work contributes a data-driven methodology for enhancing urban thermal resilience in planning and design workflows.