Tree cover mapping using hybrid fuzzy c-means and Landsat satellite images
Countrywide up-to-date tree cover maps provide valuable information for planning and management purposes to investigate location of the resources and to identify afforestation and deforestation patterns. Landsat imagery for more than three decades can provide the bases for tree cover map calculation using satellite image classification, however, practical use of classification methods is limited due to lack of user-friendly solutions and complex interpretation of the results. The objective of this study is to evaluate user-friendly hybrid classification scheme for tree cover mapping in Latvia and to explore the nature of the spectral classes and consistency of the results when methodology is applied to images of different dates. Tree cover is estimated using unsupervised fuzzy c-means with the stability check to ensure the presence of the same spectral classes in independent tests. Spectral classes are classified into two categories: tree cover and other by employing k-nearest neighbours. Such approach does not require high quality sample data and does not include user defined internal parameters of the algorithms (however, they can be specified if needed). The best overall accuracy achieved for year 2014 was 94.2 % with producer's accuracy 98.7 % (tree cover), 90.5 % (other land cover), user's accuracy 90.0 % (tree cover), 98.8 % (other land cover) and kappa 0.89. Consistency studies showed high impact (within 10 % of overall accuracy) of unique conditions during the image acquisition. Some of the spectral classes represent borderline case between tree cover and other land cover types like sparse young stands. Those cases are the main threat to the consistency between the results of different dates and seasons.