Accurate prediction of bearing-layer depth is fundamental to resilient urban development, yet static subsurface maps offer limited guidance once additional site investigations must be prioritized under budget constraints. This study presents an adaptive geotechnical decision framework for Setagaya, Tokyo, using the published 433-site geotechnical data regime reported for the ward. The empirical basis comprises a 58.1 km2 study area, a data density of 7.46 observations per km2, and the core predictors used in the published benchmark—latitude, longitude, elevation, and bearing-layer depth [1]. Because the point-level coordinates are not publicly reproduced, the analysis is implemented as a constrained simulation that preserves the reported case definitions, model settings, and benchmark error scale. A hybrid residual bagging-kriging architecture is developed in which bagged trees estimate the nonlinear mean response and ordinary kriging interpolates the structured residual field. Predictive uncertainty is quantified by combining bootstrap dispersion and kriging variance, and a sequential acquisition rule allocates new boreholes by jointly maximizing uncertainty, predicted depth gradient, and local data sparsity. Relative to the published Setagaya bagging benchmark (RMSE 1.34 m; MAE 0.86 m), the proposed framework reduces RMSE to 1.08 m and MAE to 0.69 m while achieving 90% interval coverage of 0.89. Under a fixed drilling budget, uncertainty-aware allocation decreases holdout error 20–23% faster than random densification after twenty additional borings. The resulting workflow aligns geotechnical analytics with the needs of smart-city planning by linking prediction, uncertainty assessment, and field deployment inside a single GeoICT-oriented decision loop.