From Pixels to Patterns: TerraMind at Meter Resolution with Weather Context

We present an extension of TerraMind to ingest high-resolution National Agricultural Imagery Program (NAIP) data, enabling detailed surface monitoring at meter-scale resolution. To adapt to the increased spatial granularity, we introduce a multi-scale tokenization approach that captures local and contextual features across varying patch sizes. In addition, we incorporate surface-level temperature data as a parallel input stream by tokenizing meteorological observations and aligning them with spatial tokens. This fusion allows the model to learn interactions between land surface conditions and weather dynamics at scale.

The approach was applied across diverse landscapes to support tasks such as land cover classification and crop yield predictions. Qualitative results show improved spatial precision and sensitivity to localized weather scenarios, particularly in heterogeneous regions. This work demonstrates the scalability of TerraMind to high-resolution data and highlights the benefits of structured multi-modal token ingestion for Earth surface understanding.