Testing two data fusion methods for multiscale and multiclass land-use/land-cover maps to improve fractional information at medium resolution

Barrasso, C. (2021). Testing two data fusion methods for multiscale and multiclass land-use/land-cover maps to improve fractional information at medium resolution. IIASA YSSP Report. Laxenburg, Austria: IIASA

Preview

Text YSSP_Report Barrasso_corrected.pdf - Accepted Version Available under License Creative Commons Attribution Non-commercial. Download (1MB) | Preview

Abstract

High uncertainty is found during inter-comparison of land-use/land-cover (LULC) maps derived from remote sensing imagery. Among the reasons for classification mismatch, especially in coarse maps and heterogeneous areas characterized by mixed pixels, is that the landscape heterogeneity is ignored by providing only the LULC class covering the largest portion of a pixel. Pixels are arbitrary spatial units determined mainly by the sensor’s properties and can have little relation to natural units on the ground. In fact, the use of class proportions in ground-truth training data, that better depict reality, proved to decrease the thematic accuracy of traditional LULC maps characterized by one LULC class per pixel. Because high-resolution LULC maps upscaled to coarser resolutions provide higher accuracy than natively-coarse maps, and because, except from creating new maps, integration of available ones can increase the final accuracy, during this project the potential of two data fusion methods for multi-scale (from high to coarse resolution) and multi-class maps to derive more accurate ones with fraction information at medium resolution (100m) was explored. Two data fusion models were tested in four study areas characterized by both mixed and pure-pixels by using seven LULC maps as input and a ground-truth sub-pixel database as response variable. The models’ output was then validated and compared against each individual input map, in both mixed and pure-pixels, by using the sub-pixel thematic accuracy matrix. To make more robust predictions and better answer the research questions of the study improvement of the goodness of fit of the data fusion models is needed. Despite the need of the models’ amelioration, it was observed that multiscale and multiclass data fusion improved the sub-pixel accuracy of some LULC classes compared to some of the maps used as input specially in mixed-pixels.

Actions (login required)

View Item