Simulation of subtropical courtyard microclimate evolution and optimization algorithm for landscape design based on multi-source remote sensing data fusion

Abstract Subtropical courtyards temper heat and admit light, yet their thermal performance has been hard to quantify at a scale fine enough to guide design. This study builds an end-to-end pipeline, resting on freely distributed satellite products and open code, that fuses multi-source remote sensing, simulates courtyard microclimate evolution, and optimises landscape configuration; it is applied to twenty-four courtyards in Guangzhou, southern China, a humid subtropical (Köppen Cfa) city, observed between 2021 and 2024. Complementary open products—Landsat-8/9, Sentinel-1/2, MODIS and SRTM/ASTER elevation—were harmonised through a weighted wavelet spatiotemporal fusion whose variance-driven weights marry Sentinel-2 spatial detail to MODIS temporal density, with residual-conserved downscaling holding the land-surface temperature field to its calibrated coarse-scale mean. Because that constraint alone cannot demonstrate microscale accuracy, the nominal 10 m thermal product was tested against twelve campaigns of UAV thermography: it reached a spatial RMSE of 1.62 °C and outperformed bicubic interpolation, vegetation-index sharpening, STARFM and ESTARFM, while ablation identified residual conservation as the component that matters most. From the fused fields a seven-parameter feature system was derived, and correlation screening isolated vegetation cover and impervious fraction as the dominant thermal controls. A hybrid simulator then coupled a physics-constrained convolutional–recurrent network to a ventilation-weighted cellular automaton. Validated separately for surface temperature, air temperature, humidity, mean radiant temperature and physiological equivalent temperature under a leakage-controlled spatiotemporal hold-out, it reproduced air temperature with R 2 of 0.87–0.92 and RMSE near 1.2–1.6 °C, humid-season humidity remaining the softer spot. Model development drew on eighteen courtyards and 11 916 twelve-step training sequences against 5 850 validation and 5 868 test sequences, with six further courtyards and the twelve thermography flights held outside every calibration step; the learner carries 95 778 trainable parameters, and training, test and transfer errors of 1.11, 1.26 and 1.74 degrees Celsius bracket the cost of moving to unseen geometry. Wrapping this evaluator inside a three-objective optimisation—cooling, comfort, greenery—an improved evolutionary algorithm with adaptive weighting and archive elitism attained a larger hypervolume than NSGA-II, NSGA-III, MOEA/D and SPEA2 across thirty independent runs, and needed roughly two fifths fewer simulator calls than the standard elitist baseline. Applied to one representative paved courtyard, the balanced scheme was predicted to lower mean floor temperature by 3.8 °C (95% CI 2.6–5.0 °C) and to raise in-band comfort hours by 27% points (95% CI 17–37), with relief tracking connected planting; these are model-conditioned predictions, not measured outcomes. The work delivers a transparent, auditable route from open imagery to quantitatively comparable subtropical courtyard design.