The integration of artificial intelligence into Earth Observation (EO) has long been hindered by the difficulty of connecting abstract linguistic reasoning with precise visual coordinates. While traditional vision-language models can describe a satellite image, they often struggle to pinpoint the exact pixels responsible for their conclusions. TerraScope addresses this gap by introducing a unified architecture capable of pixel-grounded reasoning, allowing the model to "show" its work by highlighting specific areas of interest on a map or satellite feed.

What sets TerraScope apart is its inherent flexibility regarding data types and timeframes. It can seamlessly process single-modality inputs like standard optical imagery or Synthetic Aperture Radar (SAR) data—which uses radio waves to map the ground—a capability crucial for monitoring regions obscured by heavy cloud cover. Furthermore, the model excels at multi-temporal reasoning, enabling it to analyze sequences of images over time to detect environmental changes or urban development patterns with high granularity.

To support this advancement, researchers curated Terra-CoT, a massive dataset featuring one million samples that utilize chain-of-thought (CoT) reasoning. By embedding pixel-level masks directly into these logical steps, the model learns to justify its spatial decisions through visual evidence. Evaluation on the new TerraScope-Bench confirms that this approach not only improves answer accuracy but also provides interpretable results, making these AI insights significantly more reliable for researchers and urban planners.

The integration of artificial intelligence into Earth Observation (EO) has long been hindered by the difficulty of connecting abstract linguistic reasoning with precise visual coordinates. While traditional vision-language models can describe a satellite image, they often struggle to pinpoint the exact pixels responsible for their conclusions. TerraScope addresses this gap by introducing a unified architecture capable of pixel-grounded reasoning, allowing the model to "show" its work by highlighting specific areas of interest on a map or satellite feed.

What sets TerraScope apart is its inherent flexibility regarding data types and timeframes. It can seamlessly process single-modality inputs like standard optical imagery or Synthetic Aperture Radar (SAR) data—which uses radio waves to map the ground—a capability crucial for monitoring regions obscured by heavy cloud cover. Furthermore, the model excels at multi-temporal reasoning, enabling it to analyze sequences of images over time to detect environmental changes or urban development patterns with high granularity.

To support this advancement, researchers curated Terra-CoT, a massive dataset featuring one million samples that utilize chain-of-thought (CoT) reasoning. By embedding pixel-level masks directly into these logical steps, the model learns to justify its spatial decisions through visual evidence. Evaluation on the new TerraScope-Bench confirms that this approach not only improves answer accuracy but also provides interpretable results, making these AI insights significantly more reliable for researchers and urban planners.