Floods are increasing both in frequency and magnitude across different parts of the world. Although technological advances have improved flood forecasting, our understanding of how floods affect communities remains limited. This study develops a multi-stage framework, integrating Natural Language Processing and the Bidirectional Encoder Representations from Transformers (BERT) approach, to identify, extract, and classify local-scale flood impacts from unstructured news articles. The flood impacts are classified into 16 impact categories at the super neighborhoods (SNs) scale in Houston, Texas. A Houston-specific gazetteer is prepared which, along with fuzzy matching, geolocates impacts from news text. The BERT model achieved an F1 score of 0.87, a precision of 0.85, and a recall of 0.9, highlighting strong and reliable performance. Results show that around one-third of SNs contribute to 90% of all reported flood impacts across Houston. Moreover, flood impacts are spatially clustered, with central, western, and southwestern regions showing disproportionate exposure. Meyerland is the most impacted SN, followed by Kingwood and Downtown. Regular Business Loss emerged as the most frequently mentioned impact category in news articles, affecting nearly 36% of SNs. Furthermore, the BERT model demonstrated strong capability in capturing the progression of Hurricane Harvey. Mentions of Evacuation rose sharply in news articles immediately after Harvey made landfall. Flood impact mentions spiked in May, July, and September, according to trends observed in the news articles. The study effectively captures fine-resolution, sector-oriented disaster impacts from text-based sources necessary for timely decision-making, impact prediction, flood risk mitigation and resilience planning….Read more