Feature Extraction from GIS images

This repository attempts to extract semantic information / features from GIS images.

Problem statement analysis and proposed solution

EDA

• The contains 3 high resolution images of sub-urban landscapes.
• The images have identifyable features like
  • Roads (Many are occluded due to presence of trees)
    • Muddy roads
    • Thar roads
  • Buildings
  • Trees
  • Barren lands
    • empty grounds
    • harvested agricutural fields in one case.
  • Water bodies
    • Pond/ Lakes
    • Ocean

We limited our study to building and roads.

Search for Existing solutions:

• We explored two github projects for existing possible pretrained models. None of them provided any reliable models.

  • Repo 1
  • Repo 2

• Kaggle - We found training scripts to train segmentation model for building & training.

Training our own segmentation

• With the training script we train two segmentation models one for buildings and another for roads.

• The same model can be extended to detect more objects like water bodies, barren lands, vehicles given sufficient data.

• The model were trained for limited epochs, due to gpu contraints however this was sufficient for the road model to catch on however the models achieved reasonable levels on performence.

Results:

• The road model is able to identify thar roads as the massachussets data shows, and identifies the roads only that are gray in appearance and misses the muddy ones.

• We can trick the model into identifying the muddy roads by twaeking image attributes like contrast and saturation.

• The building model is able to identify though the segments are not very accurate.However it is able to distintly tell apart buildings from other features like trees and and roads.

• We finally present an overlay in the results notebook

The way ahead

• To Identify more features we need to put in effort in data collection, by searching for more existing datasets and adapting them to our problem , or by manual labelling.

• We could use the pseudo label from these to extend our datasets.

• To classify terrains:

  • We could set heuristics like percentage of builing and roads cover to identify urban and sub urban space. Or percentage of forest cover to idenfy jungles.

  • Finetune explicit U-net models for this purpose.

our custom labelled data : link