iota2 workflow for SWH imagery
Spot World Heritage is a CNES program whose objective is the free availability for non-commercial use of orthorectified products derived from multispectral images of more than 5 years old from the Spot 1-5 satellites family. http://www.theia-land.fr/en/products/spot-world-heritage
Images are delivered in Level 1C (TOA ortho-images). There are several reasons why building TOC reflectance image time series with SWH data is complex using the MACCS/MAJA approach:
- images can have different acquisition angles and have a limited footprint (60 km x 60 km)
- there is no consistent archive and a particular area can be observed from 0 times to more than 10 each year (not counting cloud cover
Therefore, the land cover mapping with iota2 needs to be done using another approach. Instead of performing a classification on time series, a classification for each scene and a post-classification fusion will be implemented. Only the available pixels will be classified, since no gapfilling can be applied. Also, only the scenes overlapping training data can be classified, since a specific training for each scene has to be performed (strategies using domain adaptation to use classifiers trained on scenes which are close in time and space can be implemented). The fusion of all the classifications intersecting a particular pixel will be performed using Dempster-Shaffer's rule exploiting the confusion matrices of the individual classifications (or other fusion approaches exploiting uncertainty if/when they are available in OTB).
The steps can be summarized as follows:
- For every scene intersecting the area of interest and having training data (in-memory pipeline): 1.1 TOC to TOA conversion using generic parameters or aeronet data if available 1.2 Compute features (NDVO, NDTI, Brightness, color index) 1.3 Compute a validity mask (borders, maybe clouds) 1.4 Train a classifier using also the validity mask
- Define a unique tile grid for all the area of interest
- For every tile in the grid 3.1 For every available scene which intersects the tile (in-memory pipeline) 3.1.1 TOC to TOA conversion using generic parameters or aeronet data if available 3.1.2 Compute features (NDVO, NDTI, Brightness, color index) 3.1.3 Compute a validity mask (borders, maybe clouds) 3.1.4 Find the classifier which is the closest in time and space and apply it using the mask (invalid pixels are set to no data) and compute a confusion matrix using the validation data set 3.2 Fusion of all classifications for the tile
- Mosaic of the classifications for each tile
- Spatial regularization
- Final validation