2017–2018 Land Cover Map of Pyrénées-Atlantiques

This archive contains:

├── classification_dpt64_16_classes.tif : the 16 classes land cover map

├── classification_dpt64_16_classes_confusion_matrix.png : the confusion matrix. Have a look at it, it is performed on a different dataset than the one used for training the classifier.

├── classification_dpt64_21_classes.tif : the 21 classes land cover map including post-treatments (https://framagit.org/Schwaab/projet_predateurs64/-/blob/main/scripts/ClassificationPostProcess.py)

├── colorFile.txt : color file for symbology

├── configfile_iota2.cfg : iota2 configuration file (in case you are already using iota2. If not, what are you waiting for ?)

├── document_methodologique.pdf : technical report (french) for the classification 

├── nomenclature.txt : nomenclature file

├── reference_data_2018.shp : the training and validation data set in its 2018 version (for crops)

├── reference_data_2019.shp : the training and validation data set in its 2019 version (for crops)

├── reference_photo_interpretation.shp : the part of the training and validation data set that has been photo interpreted with a field giving the potential species or combinations of associated vegetation

├── reference_tree_nomenclature.png : a visual about the reference data

├── stratification_3_zones.shp : the stratification layer that has helped improve classification results. It is based on landscape entities (https://data.le64.fr/explore/dataset/entite-paysagere/) 

├── style_16_classes.qml : the Qgis style layer 16 classes

└── style_21_classes.qml : the Qgis style layer 21 classes
 

Description:

The land cover map of the French department Pyrénées-Atlantiques (64) is based on Sentinel-2 (L2A level) satellite images performed with Iota² chain (https://framagit.org/iota2-project/iota2/). The algorithm used is Random Forest. The time series used ranges from 2017 to 2018.

During the development phase of this classification, the collection of additional training data on the photo-interpreted classes 'landes basses' (low heath shrublands), 'landes hautes' (high heath shrublands) and 'landes hautes avec arbres' (high heath shrublands with young-growth forest) has led to a remarkable increase of the number of pixels of these classes and with it the visual quality of the map. However, this increase has been linked with only minor to almost no significant improvement of the F-scores on these classes. Some are still massively confused with other land covers like grasslands and broadleaf mature forests. Especially the mixed class 'landes hautes avec arbres' (high heath shrublands with young-growth forest).

We take it as a limit of the reference data that is built from divers data sources and would always beneficiate from more training samples of shrubby classes and a better precision of the class 'forêt de feuillus' (broadleaf mature forests). But this could also show the limit of pixel-oriented classifications for mixed/textured classes (classes with high intra-class heterogeneity). Experimentations using a contextual method – the Auto-context method now being included in Iota2 thanks to Dawa Derksen and Iota2 developers (http://lannister.ups-tlse.fr/oso/donneeswww_TheiaOSO/iota2_documentation/develop/autoContext.html) – has unfortunately not been conclusive on that matter yet.