Use of intra-annual satellite imagery time-series for land cover characterization purposes

摘要

Automatic image classification often fails at separating a large number of land cover classes that punctually may present similar spectral reflectances. To improve the classification accuracy of such situations, multi-temporal satellite data has proven valuable auxiliary information. In this paper, we present a study exploring the use fulness of intra-annual satellite images time-series for automatic land cover classification. The reported work aims at producing a land cover classification of continental Portugal from multi-spectral and multi-temporal MODIS satellite images acquired at a 500-meter spatial resolution for the year 2000. We started our study by performing a single date classification to define the month with the best score as a benchmark to compare with classification accuracies obtained with sets of images from various dates. Then, we considered various combinations of twelve intra-annual image observations (one per month) to quantify the gain when integrating temporal information in the classification process. Curiously, the results we obtained show that multi-temporal information does not significantly improve overall classification accuracy, but in particular it permits to discern between land cover classes that share similar reflectance spectra while different from their phenology (indiscernible from single date observations). Surprisingly also, we show that only few (typically 2) dates are sufficient to reach optimal performance of our multi-temporal classifier. In our study we used a Support Vector Machine learning approach.