I am interested in delineatig riparian vegetation of Danube river in Serbia, using Sentinel-2 L2A images with atmospheric corrections.
My question is regarding the best methodology for accurate delineation using this type of data.
Is SNAP Toolbox enough for all needed operation or just for preprocessing?
Do you suggest using vegetation indices (NDVI,NDWI), PCA pixel analysis or some unsupervised classification (K-means) or supervised classification (Random Forest)?
Beside this, riparian zone in this region is characterized by presence of large number of agricultural area, so one of my goals is also to classify crop areas within riparian zone.
So what are your thoughts, Multi-temporal analysis or Single Satelite image?
Vegatation indices or pixel based analysis, or combined?
Supervised or unsupervised classification?
Sentinel-2 data or combining with other Sentinel products (S1)?
Study the vegetation first. What are the typical growth cycles? For agricultural crops you will need to cover the vegetation period with repeated optical images. The classifier you choosing depends on the desired degree of detail (don’t go to complicated). NDVIs may also serve as an important
Study data availability second: If the number of available images is too low due to cloud coverage, check for radar. The calibrated amplitudes as well as the coherence change over time will serve as valuable features for your classification.
Study other available data sources than radar or optical: Global DEMs
For natural vegetation along the Danube, you may also want to consider using radar derived DSM (digital surface model, top of canopy), which often nicely show medium to tall vegetation.
Thank you for your answer and comments.
My main goal is to delineate riparian zone (80% poplar plantations, 20% meadows) from the agricultural and urban area. I’m not interested in crop classification, just identifying agricultural area inside riparian zone and around it.
So far I used 5 Sentinel-2 images form 2017. with incorporated atmospheric correction and low cloud coverage, and basically tried different delineation models (RF on reprojected images, or NDWI; NDVI, same as for K-means classification) which gave me similar results, depending on the number of trees and cluster.
Since I’m a beginner in SNAP Toolbox and satellite imagery processing in general, I’m not familiar with all possibilities and techniques necessary for correct image analysis, so every advice is precisous.