Urban Microclimate Estimation using CNN with Environmental Image Channels and Meteorological Covariates

Mitigating urban heat requires high-resolution estimation at the block–building scale; however, conventional interpolation approaches depend heavily on sensor layout, and tabular encodings of features discard spatial context such as boundaries and connectivity. We propose a non-interpolative patch-regression CNN that treats land use/land cover (LULC) and shadow as multi-channel images and fuses them with a small set of external meteorological covariates to estimate near-surface air temperature. Fixed and mobile observations are used solely as supervision during training, so inference requires no local sensor observations. In an evaluation over Tohoku University Aobayama (2m resolution; time-independent inputs), the proposed model achieved MAE comparable to distance-based Random Forest (RF) and Gradient Boosting (GBR), while delivering consistently higher R2(up to 52%). Predicted temperature maps captured plausible spatial variability aligned with shadows and land use. Moreover, ablation studies indicate that the shadow and LULC channels, as well as incorporating mobile observations during training, each contribute to improved explanatory power.