Seeing Neglect: Measuring the Predictive Value of Public Aerial and Street-Level Imagery for Property Code Violations

We measure how well freely available imagery (county and USDA aerial orthophotos and Google Street View) predicts open residential code-enforcement violations, scoring hand-crafted color/texture features, zero-shot and aerial-finetuned CLIP, a DINOv2 probe, an open-vocabulary detector (OWLv2), and a small vision-language model (Qwen2-VL-2B) against matched same-neighborhood controls in Hillsborough County, Florida. We report four findings. First, a label-free Street View detector transfers across metros: with no retraining, pooled AUC in Broward County matches Hillsborough (0.58/0.59) and overgrowth replicates (0.63/0.64). Imagery is nearly universal while code-enforcement records are not, so a detector that transfers without local labels is the one that can actually be deployed. In a third metro with dated records (Jacksonville), the signal predates the record: pre-citation Street View predicts next-year first citations (overgrowth AUC 0.61, n=1,198) with significant lead up to 12 months, although it cannot predict citation timing among eventually-cited addresses. Second, model scale is not the lever: a four-number greenness statistic outperforms every deep aerial embedding, and a holistic neglect prompt outperforms targeted object detection. Third, cues are viewpoint-specific: overgrowth is detectable from the air (AUC 0.65) while debris is visible only from the street (0.50 vs 0.62), and the 0.6-0.7 AUC ceiling is consistent with a simple label-noise attenuation model. Fourth, the viewpoints are complementary: fusing them raises every signal, although the gains are not yet significant at n=352. We release an anonymized benchmark, code, and per-parcel features.