This project is no longer maintained and will not receive any further updates. If you plan to continue using it, please be aware that future security issues will not be addressed.
An Instance Segmentation model to classify different landcover classes using raw satellite imagery.
Input: of Landsat 8 images Level 2 collection 2 using https://earthexplorer.usgs.gov/ web interface.
The Multi-spectral Image consists of Blue, Green, Red, NIR, SWIR 1 and SWIR 2 corresponding to bands numbers (2, 3, 4, 5, 6, 7), respectively.
Label: Landcover 2015 from National Forest Information System
The following classes were included in label data:
- no_change
- water
- snow_ice
- rock_rubble
- exposed_barren_land
- bryoids
- shrubland
- wetland
- wetlandtreed
- herbs
- coniferous
- broadleaf
- mixedwood
The preparation of the train data consists of extracting pairs of input und output of the train and label data. This requires the datasets to be projected in the same spatial reference. Therefore, the landsat images were reprojected to match the same spatial reference of landcover dataset. After Datasets-registration patches with fixed size were extracted to prepare the train and label data.
The model has u-net architecture consisting of 5 convolution and deconvolution layers. The model is trained to classify 4 different classes (water, herbs, coniferous and other) using the dice coefficient to evaluate accuracy. The model has reached total accuracy of 89% after learning for 120 epochs.
After the model loads the weights it can estimate raw bands images of landsat 8 using model.estimate_raw_landsat(path) as demonstrated in test.py.
The raw landsat bands should be in one folder named as their originial Landsat Product Identifier L2 followed by the SR_B<band_number>.TIF (e.g. LC08_L2SP_196024_20210330_20210409_02_T1_SR_B4.TIF is band 4 of the landsat product LC08_L2SP_196024_20210330_20210409_02_T1)
The result classified_landcover.tiff is saved as a geo-referenced one-band GeoTiff in the same folder.