As discussed many times before, we need to know the preprocessing method used to generate the 10 spectral features ['red', 'nir', 'swir16', 'swir22', 'blue', 'green','rededge1', 'rededge2', 'rededge3', 'nir08'] in the test.csv file (dividing pixel value by 10.000 as some people suggested doesn’t seem reasonable because the value range doesn’t match). The fact that the logloss score is greater than 0.7 that mean the current model is performing very bad. @Zindi, could you please provide this information to us?