Deep Learning Satellite Imagery Land Cover Classification with U-Net

Hi Geospatial Enthusiast!

From the previous video/scripts where I challenge myself using Python to classify land cover as alternative to Earth Engine. I will step it up a notch where Earth Engine could not achieve.

Repository: https://github.com/ramiqcom/deep-lear...

So I develop a code to do a deep learning land cover classification using U-Net model. I develop the model using Keras/Tensorflow library in Python.

Difference than the previous model based on pixel-based/tabular classification, we are now using spatial dimension or convolutional where it learn not only the pattern of pixel but its surrounding too. Read more about U-Net here https://arxiv.org/abs/1505.04597.

Basically what it does is that it segment the predictors (Landsat imagery) then compared it to the label (Land Cover Patches) using encoding and decoding block.

The repository consisted of two script: preprocess and modelling. Preprocess is the script to generate Landsat composite and land cover label patches. This patches then used on the second script where the modelling happen.

The modelling script also show the asssessment of the model to predict testing dataset using confusion matrix. I also apply the model to the bigger image to see the difference between the true land cover label of the full image.

You can change the parameter suits your need, such as the size of the patches, the neuron/layers size, the amount of encoding/decoding block, kernel size, etc.. Remember that it might adjust the model performance and training time!

You can use and modify this script for your project, research etc.. just don''t forget to credits me! Thanks.