Microclimate air temperature varies substantially over short distances because of local factors such as buildings, vegetation, solar radiation, and shading. Dense observations can capture this variability, but long-term deployment at many target sites is costly. This study evaluates an long short-term memory (LSTM)-based framework for air-temperature prediction at data-scarce microclimate sites using a nearby public meteorological station as a reference. The model predicts the temperature difference between the target site and the reference station instead of directly predicting the target-site temperature. Transfer learning is also introduced by pretraining the model on other nearby microclimate sites and fine-tuning it using limited target-site data. The framework was evaluated using observations from Wakayama University and the automated meteorological data acquisition system (AMeDAS) Wakayama station. Difference prediction reduced root mean square error (RMSE) compared with direct prediction at all evaluated sites. Transfer learning did not always improve the best accuracy when sufficient target-site data were available, but it improved robustness when fine-tuning data were limited or seasonally different from the evaluation period. These results suggest that difference prediction and transfer learning can support practical air-temperature prediction at data-scarce microclimate sites under conditions similar to those examined in this study.