Skip to main content

Table 4 Model evaluation metrics for the Nigeria and Guatemala test sets

From: Residential scene classification for gridded population sampling in developing countries using deep convolutional neural networks on satellite imagery

Model Type Acc. Prec. Recall F1
Nigeria
 Baseline CNN Deep 88.9% 89.2% 88.9% 89.0%
 VGG16 with ImageNet weights Deep 93.4% 93.4% 93.4% 93.3%
 InceptionV3 with ImageNet weights Deep 93.6% 93.6% 93.6% 93.6%
 VGG16 and InceptionV3 ensemble Deep 94.5% 94.5% 94.5% 94.5%
 Decision Tree Shallow 80.3% 80.9% 80.3% 78.9%
 Gradient Boosting Shallow 80.3% 80.9% 80.3% 79.0%
 AdaBoost Shallow 80.6% 81.8% 80.6% 79.2%
 Random forest Shallow 80.1% 80.7% 80.1% 78.8%
 Logistic regression Shallow 80.6% 81.8% 80.6% 79.2%
 Support vector machine Shallow 79.9% 81.5% 79.9% 78.1%
 K-nearest neighbors Shallow 75.6% 81.3% 75.6% 71.3%
 Human benchmark Human 91.0%*
Guatemala
 Baseline CNN Deep 93.3% 93.3% 93.3% 93.3%
 VGG16 with ImageNet weights Deep 96.4% 96.7% 96.4% 96.5%
 Inception V3 with ImageNet weights Deep 95.6% 95.9% 95.6% 95.6%
 VGG16 and InceptionV3 ensemble Deep 96.4% 96.7% 96.4% 96.5%
 Decision tree Shallow 93.8% 94.1% 93.8% 93.8%
 Gradient boosting Shallow 93.8% 94.1% 93.8% 93.8%
 AdaBoost Shallow 92.9% 93.1% 92.9% 93.0%
 Random forest Shallow 93.8% 94.1% 93.8% 93.8%
 Logistic regression Shallow 93.8% 94.1% 93.8% 93.8%
 Support vector machine Shallow 93.8% 94.6% 93.8% 93.9%
 K-nearest neighbors Shallow 92.4% 93.7% 92.4% 92.6%
Human benchmark Human 97.1%*
  1. *Raw agreement between two independent coders