Sidenote

In this dashboard, we have three section of contents:

• “About” is an introductory section as you see here.
• “Train” is our machine learning training results. Each square is an image tile in Liberia, and it is color-coded to show model prediction and accuracy.
  
• “Test”. It requires more than 1.05 million image tiles cover the whole of Liberia (at zoom level 17 ~ 1.192m resolution).To test out our model performance, we created 4460 test tiles that cover arbitrary areas, as shown on the map in “Test”. “Predicted School” are the results which our model considers as a school and “Provided School” are the known image tiles that contain schools.
  

Behind the model

Schools in urban and rural areas look very different in Liberia. After scrutinizing the schools in urban areas, we concluded that we should exclude them from our machine learning algorithm training (please refer to the right side “Others” image tiles). To train an algorithm on schools in rural areas was also non-trivial as it contains three elements: green forest, bare land and building(s). Our machine learning will be used to recognize schools from other not-school in term of their differences among image patterns, image content, colors and so on.

We have 2306 schools locations provided by Project Connect team, and we selected 1726 schools at rural areas to train our machine learning model. It’s a multi-label image classification built on a convolutional neural net. We used five labels for each image tile for model training: school, urban, rural, cloudy, and water. For example, for a school in the rural area of Liberia beside the water with clouds when the satellite image was taken, we will have values of 1, 0, 1, 1, 1.

Our current model performance and findings:
- 97% accuracy . (It’s from (1701 +430)/(25+34+430+1701)). Please see the bar graph in “Train”.
- 93% of validation/test accuracy . It means 2190 training images tiles were shuffled and split into 1752 (80%) as training dataset and 438 (20%) as test dataset to evaluate the model accuracy. 93% accuracy means we have 93% rate of prediction for a school image tile to be predicted as a school image tile in our 20% test dataset.

Why multi-label image classification?

Applying a convolutional neural network (known as deep learning) on image classification is a set of computer vision challenges. It classifies images to a single category, for example, if the main content of the image is cat or dog. But multi-label image classification will allow us to classify an image to more than a single label, for example, we want to know if an image has a car in it, and if so, has this car has gone through an accident. In this multi-label image classification of school searching, we want to know if an image tile contains buildings, and if so, whether these (or this) building(s) belongs to a school.

Conclusion
Our model performed unevenly from location to location. A poor prediction of a location might be caused by (1) training image tiles (with schools) are too clustered to certain areas; (2) we did not catch all the landscape variation in the pixel level (schools look very inconsistency in the tile), which means model could overfit the common-look schools, but underfit some rare-look ones. Our model could recognize schools but also mispredict a tile with bare land image tiles without any buildings and an image tile has the large building(s) as schools.

Training

Confusion matrix

Provided Schools from Project Connect Team