Synthetic aperture radar for a crop information system, a multipolarization and multitemporal approach

Abstract

Acquisition of timely information is a critical requirement for successful management of an agricultural monitoring system. Crop identification and crop-area estimation can be done fairly successfully using satellite sensors operating in the visible and near-infrared (VIR) regions of the spectrum. However, data collection can be unreliable due to problems of cloud cover at critical stages of the growing season. The all-weather capability of synthetic aperature radar (SAR) imagery acquired from satellites provides data over large areas whenever crop information is required. At the same time, SAR is sensitive to surface roughness and should be able to provide surface information such as tillage-system characteristics. With the launch of ERS-1, the first long-duration SAR system became available. The analysis of airborne multipolarization SAR data, multitemporal ERS-1 SAR data, and their combinations with VIR data, is necessary for the development of image-analysis methodologies that can be applied to RADARSAT data for extracting agricultural crop information.

The overall objective of this research is to evaluate multipolarization airborne SAR data, multitemporal ERS-1 SAR data, and combinations of ERS-1 SAR and satellite VIR data for crop classification using non-conventional algorithms.

The study area is situated in Norwich Township, an agricultural area in Oxford County, southern Ontario, Canada. It has been selected as one of the few representative agricultural 'supersites' across Canada at which the relationships between radar data and agriculture are being studied. The major field crops are corn, soybeans, winter wheat, oats, barley, alfalfa, hay, and pasture.

Using airborne C-HH and C-HV SAR data, it was found that approaches using contextual information, texture information and per-field classification for improving agricultural crop classification proved to be effective, especially the per-field classification method. Results show that three of the four best per-field classification accuracies (K=0.91) are achieved using combinations of C-HH and C-VV SAR data. This confirms the strong potential of multipolarization data for crop classification.

The synergistic effects of multitemporal ERS-1 SAR and Landsat TM data are evaluated for crop classification using an artificial neural network (ANN) approach. The results show that the per-field approach using a feed-forward ANN significantly improves the overall classification accuracy of both single-date and multitemporal SAR data. Using the combination of TM3,4,5 and Aug. 5 SAR data, the best per-field ANN classification of 96.8% was achieved. It represents an 8.5% improvement over a single TM3,4,5 classification alone.

Using multitemporal ERS-1 SAR data acquired during the 1992 and 1993 growing seasons, the radar backscatter characteristics of crops and their underlying soils are analyzed. The SAR temporal backscatter profiles were generated for each crop type and the earliest times of the year for differentiation of individual crop types were determined. Orbital (incidence-angle) effects were also observed on all crops. The average difference between the two orbits was about 3 dB. Thus attention should be given to the local incidence-angle effects when using ERS-1 SAR data, especially when comparing fields from different scenes or different areas within the same scene.

Finally, early- and mid-season multitemporal SAR data for crop classification using sequential-masking techniques are evaluated, based on the temporal backscatter profiles. It was found that all crops studied could be identified by July 21.