I recently conducted an experiment to generate field boundaries using SAM (Segment Anything Model), and I’m excited about the results. Initially, I tried using normal Sentinel-2 images with various combinations like RGB and NDVI, but they didn't yield good results. Instead, I decided to scrape high-quality RGB images from Apple's satellite imagery (no time series data). I used an automatic mask generator with low confidence settings to create masks. After that, I did some post-processing to eliminate overlapping masks and used techniques like dilation and erosion to get clearer masks. A significant challenge has always been processing this at a large scale. My approach involved processing overlapping patches to achieve finer masks. I just learned about a radial gradient method today, which is new to me. Nevertheless, I am pleased with the results; I even managed to convert these masks into GPKG polygon files. My current issue is that, to merge all the tiles for a country-scale map, I must load all the masks into memory. I wonder if you also use this method? I am wondering how do you avoid over segmenting unwanted areas like tree, roads etc. Exciting for the code release.