Run the project with input and output on Google Colaboratory
We ran the original model of DeepForest (without training on Boston-specific data) and tried to interpret results.
Some of the observations are reported here:
- Original model gives many false-positive predictions.
- The model misses many genuine tree crowns:
So training on a Boston-specific dataset is very important to improve the results and to get accurate and correct Tree Canopies.
According to DeepForest Literature, the prebuilt model will always improve by adding data from the target area. Our work found that the worth of carefully chosen hand-annotation on Boston specific dataset can yield enormous improvements in predictions. We envision that at least some fine-tuning of the prebuilt model for the majority of scientific applications will be worthwhile.
Details about how we did custom training on the Boston-specific dataset can be found here.
After training our custom model on Boston specific dataset, we saved the model weights, which can be used to predict canopies for our areas. Drive Link for the trained model can be found here.
Predictions from trained model:
As you can see here predictions are improved a lot and it is not giving any false predictions also it detected much better canopies compared to the original model.
If your input image is png/jpg file you can refer to this Google Colaboratory Notebook.
We were provided with Boston-specific Geotiff images (over 22k * 22k pixels). We created a single colab file; it just requires the path to the TIFF file on which we need to get predictions and a path to save smaller tiles of the original image. Ultimately, it shows the tree count on the smaller splitted tiles for the whole region on a map as a popup. A threshold can be used to decide whether we need to plant trees in the area or not.
The below image shows the result which is visible on the colab file.
A detailed implementation can be found in this Google Colaboratory file. (Only use this Colab file when you are using a Geotiff file as input.)