Detecting the annual areal extent of sugarcane crops in Queensland, Australia

Abstract

Sugarcane is an economically important crop for Queensland, Australia, that is increasingly a focus of environmental management, due to proximity to the Great Barrier Reef. This study responds to the demand for an ongoing, objective, and transparent assessment of the areal extent of sugarcane grown. I have combined Landsat and Sentinel-2 satellite imagery with time-series analysis, image-segmentation, and machine-learning, into a model that detects the annual areal extent of sugarcane in Queensland, since 2004. I then contrasted two variants of the model: one applied at the end of autumn (i.e. May), just before the annual harvest begins, and another applied earlier, in mid-summer (end of January), to help expedite the delivery of information for the current growing season. Both models relied heavily on contemporary land-use mapping, and the behaviour of enhanced vegetation index through summer, to detect accurately. The end-of-autumn model detected sugarcane with a user's accuracy of 93%, and a producer's accuracy of 98%, based on 596 random validation points. In comparison with historical published statistics, the end-of-autumn predictions of areal extent were typically in error by 7.1–10.1% of the observed mean, depending on the spatial scale of interest. The mid-summer model was slightly more volatile, typically in error by 6.8–11.0% of the observed mean. Overall, the results indicate that it is possible to detect—in near-real-time and with reasonable accuracy—the areal extent of sugarcane grown in Queensland in the summer of each year.