Tips and Tutorials \ Sentinel-2 Land Cover with Random Forest / XG Boost
An introduction to classifying Land Cover on S2 data using RF/XG-Boost
Type: Notebook
Categories: Features Extraction
Overview
ESA's new Earth Observation constellation Sentinel-2 provides images with high spatial resolution (10m), with many spectral bands, and a short revisit time of 5 days. It produces time series (in other words, a video) of multispectral images, over all continental surfaces in the world. One of the many practical applications of this imagery, is Land Cover mapping, which involves classifying the various objects that can be seen from a satellite : roads, forests, cities, fields, etc. This is a real supervised classification problem, which is used by several industrials, but also in the public domain (by the C.N.E.S. for instance). The idea is to use an entire year of data to classify each pixel. The temporal information is very useful for distinguishing between different crop types, and to mitigate the negative impact of clouds. There are a few reasons why this problem is challenging. First of all, the very large dimension of the multi-spectral time series : each time series is composed of 33 dates spread across the year, and each date is an image with 10 spectral bands. Therefore, the base feature space is already 330 dimensions. Secondly, there is a great amount of intra-class variation, due to the variety of cultural practices and climatic differences between different areas. Finally, the class nomenclature itself is quite challenging. In this session you will be working with a reduced nomenclature, with only 10 classes, but the full target nomenclature contains 17 classes.
In this practial session, you will test a few basic classification methods to try to solve this problem.
See Also
Questions ? Get help on the Forum