I am presently at a loss as to how i could perform classification of household roofing based on satellite images. I intend to use python/tensorflow, however I am a newbie to these programs.
I greatly appreciate any help with regard to this matter.
In general, you'd need to identify a target variable in your data set based on the certain context of the data and what you intend to classify. Then you'd need to deal with data quality issues and choose which model (classifier) to use.
Dear, Sorry for the delay.
If you are a starter on Pyton, don't take time with a complicate programm.
If you are able to undertaken folw. action it may be preferable.
1. Analyze the opportunity to use an application that you can build yourself.
2. If you are able to use machine learning programs and AI this shall be the best preparation to finalize your work. Explanation : a/cleaning the database b/transfert support of No identifyed or reconized files and convert them step by step c/Identify data validity and incorpored by version added classinfication. c/Test risks assessement on each steps of validation/integration in the Main data file d/Risks loses management mental file has to be done before any actions are undertaken / documentation of Identifyed problems too. e/Process on arborescence and classification done before - validation with fixing problems after.
I submit my proposal - hoop it will work. Feel free to come back! Enjoy inbetween
There are some things you can concern:
- Data: is it hyper-spectral images or non-spectral images?
- Noise: Shadow of cloud, low-resolution
- Ground-truth: how can you know the material of roof and is there any new material for classification (it is not cheap to label these data)? If not enough labeled data, you can think about active-learning:
Article Active Transfer Learning Network: A Unified Deep Joint Spect...
- Variant: there are different level of light or radiance. Invariant (abstract) feature using deep learning: fine-tuning data before classification
Conference Paper Spectral-Spatial Classification of Hyperspectral Image Using...
If you want to go for free open data you can probably try to exploit Sentinel-2 images. If you already have satellite imagery for your problems then you should start to build your training data by creating annotations in the form of polygons over your classes of interest (roof materials) and the can start training a model.
To create your training data and quickly see if your data and resolution is sufficient you can give a shot at our online platform (https://picterra.ch) and train a model directly there.
Best,
Frank
- Badges
- Science topic