Knowledge-Guided Probabilistic Forecasting of Urban Land-Surface Temperature in Three Moroccan Cities

Abstract Urban heat forecasting is important for climate adaptation, yet reliable multi-city prediction remains diffcult because short-term meteorology interacts with persis- tent urban form, this study evaluates a knowledge-guided probabilistic framework for next-day land surface temperature (LST) forecasting in Casablanca, Mar-rakech, and Tangier, a seven-day dynamic sequence and six static descriptors feed a heteroscedastic recurrent baseline, while a matched-seed warm-start resid- ual introduces prescribed directional contributions for built-up, impervious, and vegetation fractions. A sentinel-clean audit retained 365,500 valid daily targets from 440 verified cells and reproduced 24,902 canonical forecasting sequences, the point benchmark evaluates six models over five seeds on identical train, valida- tion, and test target sets (14,917/4,008/5,977), using minimum validation RMSE for checkpoint selection, the probabilistic dynamic+static baseline achieved the lowest mean test RMSE and MAE, 2.998 ± 0.043 and 2.313 ± 0.017◦ C; the proposed model ranked second in RMSE and third in MAE at 3.034±0.119 and 2.349±0.090◦ C, in the separately preserved validation-Gaussian-NLL-selected paired analysis, the 95% moving-block-bootstrap intervals for RMSE and Gaus- sian NLL included zero, whereas MAE favored the baseline by 0.0336◦ C. The structured mean residual nevertheless satisfies the prescribed directions exactly and raises complete-predictor consistency by 7.6, 10.4, and 24.7 percent- age points for built-up, impervious, and vegetation responses at the primary perturbation, calibration and hot-tail analyses do not show an overall uncertainty advantage, and transfer to an unseen city is mixed. The contribution is therefore a reproducible accuracy–physical-consistency trade-off within a probabilistic forecasting pipeline rather than a claim of predictive or uncertainty superiority.