/

Land Cover Classifier for Pléiades/SPOT

Assigns each of the input image pixels to a land cover class.


Introduction

For more information, please read the block description.

Block type: PROCESSING

Land cover classification that assigns each of the input image pixels to a discrete land cover class (water, vegetation, urban, barren land etc.).

This block is a beta version and the algorithm is trained with Pléiades/SPOT images over multiple locations. The block now supports running inference globally.

The block accepts 4 or 5 land cover classes. When selecting 4 classes, the model treats the classes Barren Land and Urban as a single class. Currently, the classification output consists of the following classes:

NumberLand cover class
1Water
2High vegetation (including trees)
3Low vegetation (including bushes and grass)
4Barren Land
5Urban (including roads and buildings)

If you would like to know more about the development of this land cover classifier and the scientific approach, we recommend reading this blog post. In addition, the model architecture we used is based on the work of Robinson et al. (2019). The source code is available publicly in this repository.

Accuracy assessment

Based on an independent test dataset, the algorithm accuracy is presented in the table below.

Accuracy metricsLand cover classifier with 4 classesLand cover classifier with 5 classes
Accuracy0.750.64
Jaccard0.220.51

Supported parameters

  • nclasses: Number of classes to infer (4 or 5)

Output format

AOI.clipped GeoTIFF format.

References

Robinson, C., Hou, L., Malkin, K., Soobitsky, R., Czawlytko, J., Dilkina, B.N., Jojic, N., 2019. Large Scale High-Resolution Land Cover Mapping With Multi-Resolution Data. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) 12718–12727. https://doi.org/10.1109/CVPR.2019.01301