Agricultural vegetation classification with SVM and polarimetric SAR data
2010
Polarimetric SAR data at L-band are known to be particularly well adapted for estimating moisture content and
roughness. However, many agricultural fields are generally covered by a short vegetation layer that hampers this
analysis. In fact, many applications of surface parameter retrieval methods using polarimetric SAR data over
agricultural sites revealed that parameters are underestimated over most of the fields covered by short vegetation
(e.g. grass, clovers, winter wheat). This bias is due to the electromagnetic contribution of the vegetation which
significantly modifies the polarimetric response. An identification of different kind of vegetation is necessary in
order to determine the feasibility to estimate soil moisture. The AgriSAR campaign, Agricultural Bio-/Geophysical
Retrievals from Frequent Repeat SAR and Optical Imaging, was conducted for ESA in 2006 in order to study the
agricultural vegetation. The multi-temporal datasets were acquired with the DLR's E-SAR sensor in Gormin
(Germany). From this campaign, many ground measurements were obtained: Leaf Area Index (LAI), wet and dry
biomass and soil moisture. Thus, using all information, eight agricultural vegetation classes could be characterized
independently of soil moisture. This paper presents this identification necessary to elaborate an original mapping
technique allowing localizing agricultural fields having a vegetation layer. A classification based on the support
vector machine (SVM) and on the analysis of polarimetric parameter behavior is developed using multi-temporal
images over fields covered by vegetation. The obtained vegetation maps allow the analysis of the temporal evolution
of plants. This classification has high product and user accuracy which are presented. The technique is shown to
perform well over the AgriSAR dataset.
Keywords:
- Correction
- Source
- Cite
- Save
- Machine Reading By IdeaReader
0
References
2
Citations
NaN
KQI