Nathan Torbick, Director at Applied GeoSolutions, is one of Harvest’s Synthetic Aperture Radar (SAR) experts, particularly with respect to the use of SAR and SAR optical-fusion for crop type mapping and crop management assessment. Dr. Torbick in conjunction with researchers from The University of Western Ontario, Agriculture and Agri-Food Canada, and Applied GeoSolutions, recently published their work in the Remote Sensing of Environment on A multi-temporal binary-tree classification using polarimetric RADARSAT-2 imagery. Their paper details how the Multi-temporal Binary-Tree Classification (MBTC) approach integrates the optimum scattering mechanism with machine learning techniques in order to distinguish crops and understand what aspect of the radar signal is driving that separation. They ultimately conclude that using the multi-temporal binary-tree classification with PolSAR satellite imagery provides high accuracy and achieves distinguishable structure classes among those that are very similar, and therefore usually difficult to distinguish.
A key driver of this effort is to help the science community and practitioners understand which features within complicated radar signals are most useful for identifying different crop types. As exponentially more data becomes available, and Big Data techniques like machine learning become more prevalent, and researchers prepare for new radar data streams (for example NISAR and RCM), understanding the physical meaning of the data for crop type mapping and management is increasingly important. These kinds of data are particularly useful in cloudy regions and act as a compliment to more traditional datasets from Landsat and MODIS optical imagery. With a better understanding of the meaning of all of the available datasets, scientists can help USDA and the international community, including those at GEOGLAM JECAM sites where this kind of data is collected systematically and standardized techniques for mapping crop types are being developed.
Read the full article for more detail on how the proposed Multi-temporal Binary-Tree Classification framework analyzes and harmonizes optimum scattering parameters and machine learning techniques in a substantial way.