Skip to main content

C-band SAR for Crop Types Classification

Crop information provided should be timely, geographically representative over the larger area, robust while improving itself with the ever changing needs, developments and the new research. Current research in SAR, which measurers intensity and scattered energy from the target surface, has recevied attention in the areas of agricultural monitoring. Its ability to work even in the presence of the clouds is the additional advantage, while the crop structures and the water presence in the vegetable canopy influence how the wavelengths are scattered. C-band along with optical imagery has resulted higher accuracy in determing crop types. The cropping systems to be classified; however, are complex due to different crop growing period, time and the types of the grown in the specific geography. It suggests the research to be focused on the *sensors*, *number of optical images required*, and *timing of receving the image*. 

A research was conducted to understand *how the C-band data can map crop types around different agro-ecosystems*, based on frameworks of JECAM (Joint Experiment for Crop Assessment and Monitoring), a branch of GEOGLAM: an initiation from G20 initiative to improve global agri. monitoring using EO data. Decision Tree (DT) and Random Forest (RF) classifier were tested and mapping accuracies were evaluated compairing with "optical only" and "optical and SAR" satellite data.

DT and RF classification:

The research used  the classification methodolgy developed by AAFC (Agriculture and Agri-Food Canada) who develops the ACI (Annual Space-Based Crop Inventory).

DT has been used in multiple cases of land used classifications which is also applied by USDA to prepare annual CDL (Cropland Data Layer). DT are non-parameteric models and works in discrete data providing classification rule sets. 

RF also being non parametric classifier has been found to be successful in many agri. monitoring cases with high classification accuracies compared to MLC (Maximum Likelihood Classification) in some cases. RF works by  creating multiples DTs where training data and predictive variables are branched in many numbers, from which individual trees are grown; each tree ending in end-nodes classes results in the one, which have the highest votes among all end-nodes.

Cloud presence decreases the accuracies of the classification of the optical imagery. For this, ACI has been built to process both C-band SAR and optical image data. 

Sites and Data Collection:

a. Sites

10 sites from JECAM countries participated in the study. Crop classes in each sites were from 3 and 10. Field data were acquired form in situ survey methods where vehicle-based survey of the agri. fields was done with recordings of the crop type and the location. The point observations were assigned to field-sized polygons. Classes were assigned to the polygon.

For USA sites, field data were derived from USDA CDL, the products which are developed and continuously refined over the last 20 years. 

b. SAR and optical data:

Sentinel-1 SAR data:  Interferometric Wide (IW) Ground Range Detected (GRD) high-resolution mode (spatial resolution: 20 m, image swath: 250 km)

RADARSAT-2 data: multiple modes 

  • Wide 2 (W2) ground range product (SGX) resolution of 20 m and 150 km swath; 
  • Wide 3 (W3) single look complex (SLC) resolution of 13.5 m × 7.7 m and 130 km swath
  • Standard mode beams (1 to 7) SLC resolution of 13.5 m × 7.7 m and 100 km swath; 
  • Fine Quad Wide (FQW) and Fine Quad (FQ) SLC products, resolution of 8 m and 50 km and 25 km swaths, respectively
Landsat 8 and Sentinel Optical Imagery: cloud free data as possible

Preprocessing steps For SAR data were, processed in Sentinel 7.0, and steps were followed as suggested in Dingle Robertson et. Al. (2020)
  • application of orbit file (Sentinel-1 data): exact sensor position and platform velocity is needed, provided by Sentinel 1 product  metdata, is applied to locate image acquisiton. Orbit state vectors are already applied to RADARSAT-2
  • speckle filtering: Gamma Maximum A Posteriori filter (11*11 m window) was applied, which assumes the speckle noise to be in  the Gamma distribution.   
  • terrain correction using an elevation derivation model (ortho-rectification): The effects of the angle and terrain are removed and the image is corrected according to the known coordinate system. 
Optical Data: The optical data were pre-processed using Senitinel 2 for Agriculture Project (SEN2AGRI) system. The pre-processed SAR and optical dadta were placed as three data stacks per site, resampled ot 20 m using bilinear functin and clipped to JECAM sites.

The optimized SAR and optical combined data stacks were created based on *the one optical and one SAR image per month of the growing season*, *best images which covered all full site spatial coverage or no clouds for optical imagery*, *band selection: six bands for the optical imagery, and VV and VH for SAR imagery*, *SAR data with the rainfall on the day were excluded*.

The DT and RF classifier were applied to each three data stacks for each JECAM site. The See 5.0 DT parameters weter set and for RF 150 trees and 10 variables were set to improve the processing time. Accuracy were accessed by the error matrix. OAs anfd  users and producers accuracies were focused. 

Results and Conclusions:
RF classifer was found to have higher accuracy. The OAs tend to fall with the decrease of the number of the SAR images. The SAR only data were not found to have more superiority over the optical data classification. The timing of image acquition is also likely to influence overall classification accuracy. Similarly, the increase in the number of the crops to be classifed also decreased the overall accuracy. 

The overall implication is the application of SAR imagery when the cloud free optical imagery is not available. The optimized combined data stack was proved to produce higher accuracies across the classes. In the cases where, the growing period is missed, the lower accuracies were achieved. 

In an attempt to study the following article:
  1. Laura Dingle Robertson, Andrew M. Davidson, Heather McNairn, Mehdi Hosseini, Scott Mitchell, Diego de Abelleyra, Santiago Verón, Guerric le Maire, Milena Plannells, Silvia Valero, Nima Ahmadian, Alisa Coffin, David Bosch, Michael H. Cosh, Bruno Basso & Nicanor Saliendra (2020) C-band synthetic aperture radar (SAR) imagery for the classification of diverse cropping systems, International Journal of Remote Sensing, 41:24, 9628-9649, DOI: 10.1080/01431161.2020.1805136


Recent Posts Widget