Mapping forest types over large areas with Landsat imagery partially affected by clouds and SLC gaps

The ecosystem services that forests provide depend on tree species composition. Therefore, it is important to map not only forest extent and its dynamics, but also composition. Open access to Landsat has resulted in considerable improvements in remote sensing methods for mapping tree species, but most approaches fail to perform when there is a shortage of clear observations. Our main goal was to map forest composition with Landsat imagery in various data availability conditions, and to investigate how the missing data, either due to clouds or scan line problems affect classification accuracy. We tested a data driven approach that is based on multi-temporal analysis of the tree species’ spectral characteristics making it applicable to regional-scale mapping even when the gap-free imagery is not available. Our study area consisted of one Landsat footprint (26/28) located in Northern Wisconsin, USA. We selected this area because of numerous tree species (23), heterogenic composition of forests where the majority of stands are mixed, and availability of high-quality reference data. We quantified how classification accuracy at the species level was affected by a) the amount of missing data due to cloud cover and Scanning Line Corrector (SLC) gaps, b) the number of acquisitions, and c) the seasonal availability of images. We applied a decision tree classifier, capable of handling missing data to both single- and a three-year Landsat-7 and Landsat-8 observations. We classified the dominant tree species in each pixel and grouped results to forest stands to match our reference data. Our results show four major findings. First, producer’s and user’s accuracies range from 46.2% to 96.2% and from 59.9% to 93.7%, respectively for the most abundant forest types in the study area (all types covering greater than 2% of the forest area). Second, all tree species were mapped with overall accuracy above 70% even in when we restricted our data set to images having gaps larger than 30% of the study area. Third, the classification accuracy improved with more acquisitions, especially when images were available for the fall, spring, and summer. Finally, producer’s accuracies for pure-stands were higher than those for mixed stands by 10 to 30 percentage points. We conclude that inclusion of Landsat imagery with missing data allows to map forest types with accuracies that previously could be achieved only for those rare years for which several gap-free images were available. The approach presented here is directly applicable to Landsat-like observations and derived products such as seasonal composites and temporal statistics that miss 30% or more of the data for any single date to develop forest composition maps that are important for both forest management and ecology