Table 1 The results of performance evaluation of traditional machine learning algorithms and DeepLabv3+ algorithm.

From: Semantic segmentation of PolSAR image data using advanced deep learning model

Parameter

RF

KNN

SVM

DeepLabv3+

Pixel accuracyPatch 1

74.74%

78.68%

74.44%

83.51%

Pixel accuracyPatch 2

81.09%

74.31%

79.59%

87.78%

Overall pixel accuracy

77.92%

76.48%

77.02%

85.65%

F1 score

0.7784

0.7762

0.7646

0.8520

Precision (Urban Class)

0.8958

0.7984

0.8977

0.9228

Training time

4204 s

41,091 s

2803 s

13411 s

Inference time (per scene)

3.04 s

864.94 s

0.46 s

0.31 s

Algorithm complexity

\(O\left( {V {\text{x }}NlogN} \right)\)b

\(O\left( {KND} \right)\)b

\(O\left( {N^{2} } \right)\)

\(O\left( N \right)\)a

  1. aIt is considered only the impact of the training set of size N, however, there are other factors (number of features, dimensionality, etc.) that can also affect the algorithm’s time complexity42,92.
  2. bIn RF, the variable V = (ntree x mtry), where ‘ntree’ is the number of trees to build and ‘mtry’ represents how many variables need to sample at each node28. In KNN, ‘D’ is the distance between the training set observation and new observation, and ‘K’ represents the number of neighbors27. The variable ‘N’ represents the number of training set size used in SVM29 and other listed algorithms.
Back to article page