The Tocantins Framework: A Machine Learning-Based Assessment of Intra-urban Thermal Anomalies

Isaque Carvalho Borges , Johari Barrientos-Murray, Lucca Pereira da Cunha, Isabela Nunes Toniolo

The Urban Heat Island (UHI) effect has been extensively studied at the city scale. Yet, the Intra-Urban Heat Island and the Intra-Urban Cool Island effects remain poorly characterized due to the absence of standardized quantification frameworks. This study introduces the Tocantins Framework, a dual-metric system combining machine learning and spatial morphology to identify and quantify intra-urban thermal anomalies in Landsat imagery. The framework consists of two metrics: the Impact Score (IS), which assesses the spatial extent and thermal intensity of Extended Anomaly Zones (EAZs), and the Severity Score (SS), which quantifies core thermal anomalousness. A two-stage detection methodology first identifies statistical temperature extremes (2nd and 98th percentiles), then applies Random Forest regression to distinguish pixels exhibiting thermally anomalous behavior unexplained by spectral indices (NDVI, NDWI, NDBI, NDBSI). The model achieved R² = 0.8023 (RMSE = 1.1612°C) when predicting Land Surface Temperature (LST) from spectral properties. We capture residual variance for anomaly detection. When applied to a Landsat 8 composite from the first semester of 2025 for Palmas, Tocantins, Brazil, the framework identified 103 thermal anomalies (76 IUHIs and 27 IUCIs) covering 30.42 km² of the urban area. Impact Scores ranged from−14.69 to 13.32 and Severity Scores from −14.49 to 16.87. The two metrics showed a weak correlation (R² = 0.154), confirming they capture complementary and distinct aspects of anomaly phenomena. The framework provides a reproducible, open-source methodology for evidence-based urban heat mitigation planning, enabling practitioners to identify priority intervention areas based on both thermal intensity and spatial impact. The complete implementation is available as a Python package via PyPI.