Jag har byggt KartGPT. En tjänst som kombinerar en stor bildmodell med en liten språkmodell för att snabbt ta fram infromation från Lantmäteriets data som geopackage. Den fungerar bara i Sverige men inte på svenska. Använd korta promts som "road, building, bare soil". www.kartgpt.se bon appetit!
Geoint AB presents KartGPT, a service that lets you extract information from high-resolution aerial imagery of Sweden using text prompts. Join our webinar (in Swedish) next Tuesday at 13:00. forms.gle/zMBvSx9suXyr...
I need one!
Ett toppenexempel på hur dagens politiker egentligen inte tror på handling, bara på att säga rätt saker.
En inventering visar att 18 procent av kulturlämningarna har en skada eller en grov skada. Markberedning och skogsmaskiner är de största bovarna i dramat.
– Förstörda kulturlämningar växer inte tillbaka, säger Michael Lehorst, kulturmiljöspecialist på Skogsstyrelsen.
Geoint is live!
Today is International GIS Day, and we’re taking the opportunity to soft-launch a service that, in just a few minutes, shows whether your idea can be realised using AI trained on laser data:
ml.geoint.se
If you are bored, you can watch my Docent lecture today at 11:00: www.slu.se/kalender/202... AI-powered maps and other problems.
Vacation from my research role gives me more time for side-projects
Don't forget to optimize the file size of the .READMEs
My landscape cubes now have a modelled groundwater level. Excited to see what my 3D-CNN can learn from this.
I’ve released a prototype of a tool that trains a machine learning model on open geographical data using your own training data.
Upload a point shapefile where the first column contains the values you want to predict.
It's free to experiment with, and no data is stored. geoint.konvergens.se
I have now added the color from IR imagery into my "landscapecubes". Now working to add groundwater under the DEM surface and then I need to figure out a way to encode logic in my 3D-CNN. For example geographical location or runoff data.
I am running some deep-learning models on voxel cubes of remote-sensing data instead of images to map soil moisture. It works but is quite demanding. So far, I have built a proof of concept in Tensorflow using 3D convolutions. Do you have any good advice for me?
When extracting streams from high-resolution digital elevation models, road culverts can be a pain. I explored an approach where a residual attention UNet was used to identify and breach road culverts, utilising ALS data and orthophotos. It works but produces a fair amount of false positives.
