Fig 1.
Land use and land cover and location of the Haean catchment.
The map on the left shows the original polygon data set by [20] (WGS84 / UTM 52N; EPSG:32652) for the year 2010 with the 14 lulc classes used in this study.
Table 1.
Distribution of the 14 LULC classes in the raster data set.
The first 6 classes were used for classification.
Fig 2.
Synthetic points (crosses denoted s1 through s5) generated by smote along the connection lines between a point (black dot denoted Qi) and its k nearest neighbours (black dots). Here, k = 5 and oversampling rate N = 5.
Fig 3.
Reflectance values of all pixels in the original data set in the red band B1, the near-infrared band B2, the blue band B3 and the mid-infrared band B7. The plain lines show the median and the shaded areas the range (from minimum to maximum value). DOY is a day of the year derived from the true acquisition date through interpolation.
Fig 4.
Jeffries–Matusita distances between different classes in the original data calculated for every modis band.
Fig 5.
Mutual information MI* between class labels and MODIS spectral bands.
Results from 10 repetitions on 6 training folds in scenarios S1 through S4: (a) red band B1, (b) near-infrared band B2, (c) blue channel B3 and (d) mid-infrared band B7. The plain lines show the median and the shaded areas the 5% to 95% quantile range.
Fig 6.
Median values from 10 repetitions in scenarios S1 (a) through S4 (d). A point on the diagonal (grey line) indicates a random guess. The order of the classes in the legend reflects the decreasing number of original pixels.
Fig 7.
Predicted land use and land cover classes.
Scenarios S1 (a) through S4 (d) and the original data set (e). The maps (WGS84 / UTM 52N; EPSG:32652) are from repetitions with the largest F-score. Classes with less than 20 original pixels and cloud contaminated pixels are marked as ‘NA’.
Fig 8.
Proportion of pixels with more than 50% of the main lulc class in the original data set.