Devlog 12: last devlog before i ship!
This was mostly bug fixing and polishing.
I finally tracked down an issue where the model was outputting 0.0. At first I thought the model itself was broken, but it turned out I was double normalizing the Google Earth Engine test data. After fixing that (and retraining the model), accuracy improved quite a bit. I also realized that green ≠ forest and overall more confidence.The permit querying system also had problems. I originally trusted gemini(clanker) because I couldn’t make sense of the IBAMA(its in portuguese) data, but that ended up sending me down a debugging forever. Even after fixing it, the ML still wasnt perfect, so i added a mask(as seen in img 3)and even after retraining the model i noticed the mask was random globs and lines, after 4 hours of trying to debug it was a few things: the model just isn’t meant for it as an u-net and i thought subtracting the two predictions would work but didn’t, so i added threshold, a bit more accurate but not good enough, missed a ton, what really fixed i t was averaging after sigamoid, binary opening and closing, and a big fix for the models accuracy(before threshold-ing) was fixing the payload, i used the wrong cords in payload so the image wasn’t correct. after all of that i finally fixed the prediction!
features i added and fixed:chain of thought, damage %, some frontend overhauls(as you can see), and more. refined prompt for agent, dynamic scanning(no longer hard coded), and chunking for memory. that’s all(all i remember at least)!