The t-SNE Algorithm as a Tool to Improve the Quality of Reference Data Used in Accurate Mapping of Heterogeneous Non-Forest Vegetation

Abstract

Supervised classification methods, used for many applications, including vegetation mapping require accurate “ground truth” to be effective. Nevertheless, it is common for the quality of this data to be poorly verified prior to it being used for the training and validation of classification models. The fact that noisy or erroneous parts of the reference dataset are not removed is usually explained by the relatively high resistance of some algorithms to errors. The objective of this study was to demonstrate the rationale for cleaning the reference dataset used for the classification of heterogeneous non-forest vegetation, and to present a workflow based on the t-distributed stochastic neighbor embedding (t-SNE) algorithm for the better integration of reference data with remote sensing data in order to improve outcomes. The proposed analysis is a new application of the t-SNE algorithm. The effectiveness of this workflow was tested by classifying three heterogeneous non-forest Natura 2000 habitats: Molinia meadows (Molinion caeruleae; code 6410), species-rich Nardus grassland (code 6230) and dry heaths (code 4030), employing two commonly used algorithms: random forest (RF) and AdaBoost (AB), which, according to the literature, differ in their resistance to errors in reference datasets. Polygons collected in the field (on-ground reference data) in 2016 and 2017, containing no intentional errors, were used as the on-ground reference dataset. The remote sensing data used in the classification were obtained in 2017 during the peak growing season by a HySpex sensor consisting of two imaging spectrometers covering spectral ranges of 0.4–0.9 μm (VNIR-1800) and 0.9–2.5 μm (SWIR-384). The on-ground reference dataset was gradually cleaned by verifying candidate polygons selected by visual interpretation of t-SNE plots. Around 40–50% of candidate polygons were ultimately found to contain errors. Altogether, 15% of reference polygons were removed. As a result, the quality of the final map, as assessed by the Kappa and F1 accuracy measures as well as by visual evaluation, was significantly improved. The global map accuracy increased by about 6% (in Kappa coefficient), relative to the baseline classification obtained using random removal of the same number of reference polygons.