You or the CNN? Land use from satellite images
You see 10 real Sentinel-2 satellite images from the EuroSAT data set. Each image covers 640 m × 640 m of land in Europe (64 × 64 pixels of 10 m). For each image, choose its land-use class among four options. After each answer you see what the convolutional neural network (CNN) of Lecture 4 predicted and the true class.
The network is a ResNet-18 (residual network with 18 layers) pre-trained on the ImageNet photographs and fine-tuned on 16,200 EuroSAT images (transfer learning). It never saw the images of the game: they come from the 5,400 test images. Two cards in each game are images that the network classifies wrongly.
What the game shows
- On all 5,400 test images the fine-tuned network is right in 98.76% of cases: 67 errors, most often annual crop taken for permanent crop (8 images).
- Transfer learning matters most when labels are scarce: with only 10 labelled images per class, the fine-tuned network reaches 87.8%, a CNN trained from scratch 69.9%.
- The game over-represents the network's errors (2 cards in 10, against 1.24% of the test images), so its score in the game is lower than on the full test set.
Data: EuroSAT, Sentinel-2 images (Helber et al., 2019; 27,000 images, 10 classes). The 40 images of the game are drawn at random from the test images with the seed 20261108: 3 per class among those the network classifies correctly and 10 of its 67 errors; every game draws 8 correct and 2 wrong ones. Model: cached predictions of lectures/L4/analysis_L4.py. Code: lectures/L4/make_game_L4.py.

