Potential of April-June multi-temporal images for crop mapping before harvest: A case study of Kashgar

This research aims at evaluating the potential of multi-temporal images acquired in April-June period for crop mapping before harvest in Kashgar China. Firstly, images of both Landsat-5 TM and Huan Jing (HJ)-1 CCD data were used to acquire an image time series with 30 m spatial resolution and 15 day temporal resolution during the entire growing season. Subsequently, period-by-period separability of different crops was measured by calculating Jeffries-Matsushita (JM) distance. Afterwards, a support vector machine (SVM) was used to compare the classification accuracy of entire growing season and April-June images. The result indicated that the late August image is the best time period to identify crops because the JM distances of pair-wise crop were larger than 1.75. And the average JM distance of April-June images was 1.939, which was slightly lower than that of entire growing season (JM distance 1.999). In addition, the overall accuracy of classification using April-June images was 85.16%. It was 8% lower than that of optimal images of entire growing season (93.74%). The misclassifications of April-June images were mainly attributed to the misclassification between wheat and wheat-maize, as summer were still in an early growing stage in late June. Overall, the research showed that domineering crops in Kashgar can be extracted one month before harvest using multi-temporal images obtained during April and June.