Due　to　the　variability　in　bi-temporal　high-resolution　images,　such　as　sensor,　shooting　angle　and　seasonal　turnover,　pseudo-change　phenomena　often　exhibited　in　urban　building　change　detection.　To　solve　this　problem,　a　new　network　model　AMDR-Net　is　proposed,　which　integrates　the　modeling　of　the　directional　relationship　between　buildings　and　their　shadows.　The　existing　modeling　methods　for　the　directional　relationship　between　buildings　and　shadows　rely　on　the　metadata　of　solar　elevation.　To　meet　this　challenge,　AMDR-Net　constructs　a　network　with　a　CNN　architecture.　Automatically　learn　the　direction　relationship　between　buildings　and　shadows　from　high-resolution　remote　sensing　images.　Experiments　conducted　on　the　Fuzhou　dataset　demonstrate　that　the　proposed　method　achieves　high　consistency　with　ground-truth　change　maps,　with　an　F1-score　of　90.83.　The　results　indicate　that　AMDR-Net　exhibits　superior　robustness　and　performance　in　mitigating　false　changes　caused　by　image　disparities,　outperforming　existing　comparative　methods.　©　2025　IEEE.