Land Cover Classification (Sentinel-2)

Land cover describes the surface of the earth. Land cover maps are useful in urban planning, resource management, change detection, agriculture, and a variety of other applications in which information related to earth surface is required. Land cover classification is a complex exercise and is hard to capture using traditional means. Deep learning models are highly capable of learning these complex semantics, giving superior results.
 

Using the model

Follow the guide to use the model. Before using this model, ensure that the supported deep learning libraries are installed. For more details, check Deep Learning Libraries Installer for ArcGIS.

 

Fine-tuning the model

This model can be fine-tuned using the Train Deep Learning Model tool. Follow the guide to fine-tune this model.
 

Input

Raster, mosaic dataset, or image service. (Preferred cell size is 10 meters.)   Note: This model is trained to work on Sentinel-2 Imagery datasets which are in WGS 1984 Web Mercator (auxiliary sphere) coordinate system (WKID 3857).   Output

Classified raster with the same classes as in Corine Land Cover (CLC) 2018. 

Applicable geographies
This model is expected to work well in Europe and the United States.
 

Model architecture This model uses the UNet model architecture implemented in ArcGIS API for Python.
  Accuracy metrics

This model has an overall accuracy of 82.41% with Level-1C imagery and 84.0% with Level-2A imagery, for CLC class level 2 classification (15 classes). The table below summarizes the precision, recall and F1-score of the model on the validation dataset.

 

ClassLevel-2A ImageryLevel-1C ImageryPrecisionRecallF1 ScorePrecisionRecallF1 ScoreUrban fabric0.810.830.820.820.840.83Industrial, commercial and transport units0.740.650.690.730.660.7Mine, dump and construction sites0.630.520.570.690.550.61Artificial, non-agricultural vegetated areas0.700.460.550.670.470.55Arable land0.860.900.880.860.890.87Permanent crops0.760.730.740.750.710.73Pastures0.750.710.730.740.710.73Heterogeneous agricultural areas0.610.560.580.620.510.56Forests0.880.930.900.880.920.9Scrub and/or herbaceous vegetation associations0.740.690.720.730.670.7Open spaces with little or no vegetation0.870.840.850.850.820.84Inland wetlands0.810.780.800.820.770.79Maritime wetlands0.740.760.750.870.890.88Inland waters0.940.920.930.940.910.92Marine waters0.980.990.980.970.980.98

 

This model has an overall accuracy of 90.79% with Level-2A imagery for CLC class level 1 classification (5 classes). The table below summarizes the precision, recall and F1-score of the model on the validation dataset.
  ClassPrecisionRecallF1 ScoreArtificial surfaces0.850.810.83Agricultural areas0.900.910.91Forest and semi natural areas0.910.920.92Wetlands0.770.700.73Water bodies0.960.970.96
  Training data This model has been trained on the Corine Land Cover (CLC) 2018 with the same Sentinel 2 scenes that were used to produce the database. Scene IDs for the imagery were available in the metadata of the dataset.

Sample results
Here are a few results from the model. To view more, see this story.