Earth Observation (EO) has seen significant growth, yet running EO models presents challenges due to the complexities of geospatial data. We have developed “EasyEarth”, a QGIS-based plugin addressing these issues by enabling users to run EO, computer vision, or vision-language directly within QGIS's familiar interface, streamlining model deployment and integration.

In particular, EasyEarth aims to overcome the major EO challenge of inefficient annotation processes. Tools such as GeoSAM have provided initial resources to facilitate the generation of training labels based on the Segment Anything Model (SAM). However, this tool supports only one pre-trained model and involves a two-step process: creating image embeddings and generating training labels using inferences with prompts. Additionally, installation requires modifying base Python libraries on QGIS, which can be insecure with potential disruption of the software environment.

To advance beyond existing annotation plugins on QGIS, EasyEarth incorporates multiple pre-trained models from popular AI communities such as HuggingFace, and streamlines and automates the generation and loading of embeddings, among other advantages. To address the potential conflicts between the model environment and QGIS, we wrap the model running environment within a Docker container and use Flask to facilitate communication between the QGIS interface and the model running environment. This separation ensures that changes in the model environment do not interfere with the main QGIS application, enhancing both security and stability.

We expect this plugin to increase the ease of using different EO models on custom data, and the efficiency and accuracy of the labeling process. The streamlined and simplified processes are expected to encourage more users, including those with limited computer science and remote sensing backgrounds, to adopt this tool in their work, facilitating broader engagement and application in various fields.