Random Forest result

snap
1

Hello,

i have problems to interpret my Random-Forest result:
RandomForest classifier newClassifier

Cross Validation
Number of classes = 2
class 0.0: 50percent
accuracy = 0.9971 precision = 0.9600 correlation = 0.9783 errorRate = 0.0029
TruePositives = 48.0000 FalsePositives = 2.0000 TrueNegatives = 638.0000 FalseNegatives = 0.0000
class 1.0: geo_50_percent_Polygon
accuracy = 0.9971 precision = 1.0000 correlation = 0.9783 errorRate = 0.0029
TruePositives = 638.0000 FalsePositives = 0.0000 TrueNegatives = 48.0000 FalseNegatives = 2.0000

Using Testing dataset, % correct predictions = 99.7093
Total samples = 1376
RMSE = 0.053916386601719206
Bias = -0.0029069767441860517

Distribution:
class 0.0: 50percent 96 (6.9767%)
class 1.0: geo_50_percent_Polygon 1280 (93.0233%)

Testing feature importance score:
Each feature is perturbed 3 times and the % correct predictions are averaged
The importance score is the original % correct prediction - average
rank 1 feature 1 : multitemporal score: tp=0.2374 accuracy=0.2374 precision=0.4752 correlation=0.7590 errorRate=-0.2374 cost=-5.0326 GainRatio = 0.1709

I get good results for the classes but bad results for the whole image.

The accuracy from 0.2374 is verry low.

Can someone explain the result?

2

Information about the metrics are nicely compiled here:
Accuracy/Precision/Correlation/TruePositive/FalsePositive/TrueNegative/FalseNegative: https://en.wikipedia.org/wiki/Sensitivity_and_specificity

RMSE, Bias and MAE: https://medium.com/human-in-a-machine-world/mae-and-rmse-which-metric-is-better-e60ac3bde13d

Low accuracy can have many reasons:

  • Not enough training samples to effectively make use of the randomization (comments here)
  • Training samples which are not representative for the classes
  • Inhomogenous training samples (one solution here)
  • Not enough feature rasters (comments here)
  • Not enough trees (comments here)
1 Like