AppliedMachineLearningPublic

I chose to predict image 9094 which was a seemingly densly populated area. The model predicted 0.78728527, however the labels was 116.1895447.

I also wanted to try and predict the label of a seemingly deserted area so I chose image 9397 which was essentially just roads without any people or buildings The model predicted 4.8975797 but the actual label was again 116.1895447. In this case I feel like my model was more accurate and I am not sure that the label is very accurate (it certainly seems like there are no people in this frame, but perhaps, as we talked about in class, the area of the label and the frame of the image do not quite match up).

Finally, I wanted to see how my model would do with an image that was somewhere in-between high-density and no-density so I chose image 9961 which had just a few buildings in its frame. The model predicted: 0.62696415 but the actual label was 99.66568756.

In some cases it seems like the labels were a bit over-zealous, but this may be because of mis-alignment between the population counts and the rasters. However, overall it seems that my model was abismal at predicting the correct labels (or even getting close to the correct value) as the prediction was often off by close to 100. This might be in part because I did not use a CNN to run my predictions. Moreover, my model was trained on a very small set of photos so increasing the number of training photos that I brought in or upsampling might be another way to make my model more accurate.