Synergy of multitemporal ERS-1 SAR and Landsat TM data for classification of agricultural crops

摘要

The objective of this research was to evaluate the synergistic effects of multitemporal European remote sensing satellite 1 (ERS-1) synthetic aperture radar (SAR) and Landsat thematic mapper (TM) data for crop classification using a per-field artificial neural network (ANN) approach. Eight crop types and conditions were identified: winter wheat, corn (good growth), corn (poor growth), soybeans (good growth), soybeans (poor growth), barley/oats, alfalfa, and pasture. With the per-field approach using a feed-forward ANN, the overall classification accuracy of three-date early- to mid-season SAR data improved almost 20%, and the best classification of a single-date (5 August) SAR image improved the overall accuracy by about 26%, in comparison to a per-pixel maximum-likelihood classifier (MLC). Both single-date and multitemporal SAR data demonstrated their abilities to discriminate certain crops in the early and mid-season; however, these overall classification accuracies (<60%) were not sufficiently high for operational crop inventory and analysis, as the single-parameter, high-incidence-angle ERS-1 SAR system does not provide sufficient differences for eight crop types and conditions. The synergy of TM3, TM4, and TM5 images acquired on 6 August and SAR data acquired on 5 August yielded the best per-field ANN classification of 96.8% (kappa coefficient = 0.96). It represents an 8.3% improvement over TM3, TM4, and TM5 classification alone and a 5% improvement over the per-pixel classification of TM and 5 August SAR data. These results clearly demonstrated that the synergy of TM and SAR data is superior to that of a single sensor and the ANN is more robust than MLC for per-field classification. The second-best classification accuracy of 95.9% was achieved using the combination of TM3, TM4, TM5, and 24 July SAR data. The combination of TM3, TM4, and TM5 images and three-date SAR data, however, only yielded an overall classification accuracy of 93.89% (kappa = 0.93), and the combination of TM3, TM4, TM5, and 15 June SAR data decreased the classification accuracy slightly (88.08%; kappa = 0.86) from that of TM alone. These results indicate that the synergy of satellite SAR and Landsat TM data can produce much better classification accuracy than that of Landsat TM alone only when careful consideration is given to the temporal compatibility of SAR and visible and infrared data.