Pixel-Level Urban Housing-Price Mapping Based on AlphaEarth Foundations: Evidence from 36 Chinese Cities

Shuyang Hou, Haoyue Jiao, Ziqi Liu, Donghai Liu, Lutong Xie, Shaowen Wu, Guanyu Chen, Yaxian Qing, Qingyang Xu, Xuefeng Guan, Huayi Wu

This study develops and validates a multi-source modelling framework for continuous, pixel-level urban housing-price mapping using surface embeddings from AlphaEarth Foundations (AEF). Pixel-level labels calibrated against multi-source market data are constructed for 288 city–year samples across 36 Chinese cities (2017–2024), and AEF’s 64-dimensional, 10 m annual surface embeddings are systematically evaluated across four training strategies, fourteen feature combinations, and four baseline models, with validation through PCA ablation, SHAP attribution, IDW comparison, spatial clustering, and temporal consistency analysis. AEF embeddings alone yield limited predictive power, with pixel-level R² of only 0.66 and 0.15 under random and spatial cross-validation, respectively; incorporating points of interest, nighttime light, population density, spatial structure, and remote-sensing indices raises these values to 0.84 and 0.49. Full-coverage prediction surfaces for 36 cities over eight years achieve a median Pearson r of 0.948; spatial clustering intensity closely matches the IDW benchmark (Moran's I = 0.865), and the median adjacent-year structural similarity index reaches 0.986. Analysis further reveals that AEF implicitly encodes substantial built-form information—explicit physical features such as building height and footprint area provide only marginal gains—yet price-relevant signals are distributed across the full 64-dimensional space without a stable cross-city subset, confirming that socioeconomic auxiliary factors remain indispensable for generalisation. Building on the continuous prediction surfaces, downstream applications—housing-price tier transitions, CBD premium gradients, and spatial redistribution during market corrections—show that surface foundation-model embeddings support not only housing-price prediction but also large-scale, continuous, and temporally consistent mapping and spatial analysis of urban economic variables.