Determining temporal windows for crop discrimination with remote sensing: a case study in south-eastern Australia

The classification of crops from remote sensing has become an important part of agricultural management, and as a result, has instigated a great deal of research aimed at increasing classification accuracy through various methods and techniques. However, comparatively little research has been performed on determining the best time(s) of image acquisition for crop discrimination even though this could impact classification accuracy as much as choice of clustering algorithm or selection of training data, for example. This case study was conducted to: (1) determine temporal windows for highest overall and individual crop discrimination; and (2) compare simple methods for combining best single-date results to increase overall accuracy. Seventeen single-date classifications of four major summer crops (rice, maize, sorghum, and soybeans) were assessed for a single growing season at the Coleambally Irrigation Area, Australia using Landsat Enhanced Thematic Mapper data. Per-pixel classifications were generated using a maximum likelihood classifier and were then combined with field boundaries to produce per-field classifications, based on the majority crop type within each field. Multi-date classifications were performed by: (1) combining various numbers of bands per date into a single image stack prior to classification (2-date, and 3-date-termed standard multi-date classification); as well as (2) extracting maximum accuracy single-crop classes from different dates and combining them, post-classification (termed iterative multi-date classification). Results showed that the general time-frame for highest overall single-date classification accuracy was late February to mid March. For the individual crops, late November to early December resulted in the highest accuracy for discriminating rice, maize was best discriminated from mid-February to mid-March, the maximum sorghum separability occurred from early April till at least early May, and the soybean temporal window extended from early January to mid-March, with late-February to early March being best. The iterative approach resulted in higher accuracy than the standard multi-date image stack of the same dates. Highest multi-date accuracy resulted from the 3-date per-field iterative classification (overall classification accuracy of 95.8%), an improvement of more than 6% over best per-field single-date results (10 March, 89.4%). Determination of temporal windows for crop discrimination and use of an iterative technique to combine multiple-date images both greatly improved overall crop classification, thus increasing the benefit of remote sensing in operational management. Crown Copyright (C) 2004 Published by Elsevier B.V. All rights reserved.