GeoAI SAM

prasanthkrishnavarman

I have used the new Text SAM (Segment Anything Model) available in the Esri India Living Atlas to detect buildings from a satellite image with 0.3m resolution using the text prompt "Buildings." However, the model only detected a few buildings. How to improve the detection accuracy of Text SAM Model? Additionally, is there a way to fine-tune the SAM model for better results?

SwastikaDas

The issue you're encountering may be due to the varying sizes of buildings in the satellite imagery. For instance, in an urban context, there are buildings of small, medium, and large footprints.  Although deep learning models are typically trained on multiple imagery resolutions, the variation in building sizes can pose challenges. To improve detection, try manually adjusting the cell size in the environments tab of Detect Objects Using Deep Learning tool between 0.1 to 0.5. This should enhance detection across all building sizes. For more details, refer to this blog post: "https://www.esri.com/arcgis-blog/products/arcgis-pro/analytics/multiresolution-object-detection-with-text-sam/?srsltid=AfmBOoohMVcSgbSg-K_J2jtp9HCdvsALOdiEeKT02Ss-d15y2ZU_Pk6L"
Regarding your second question, yes, the SAM model can be fine-tuned. The SamLoRA model, a variation of SAM now available in ArcGIS, introduces trainable layers called Low-Rank Adaptation (LoRA) to the frozen image encoder of SAM. You can train the SamLoRA model to detect specific features of interest for more accurate results.