GSOC Project

DeepForest

This project aims to make the model which would already be trained for the classification of species and detection of alive and dead, trees or birds using transfer learning on the current release model which is based on object detection, only detecting trees and birds, for now, It also involves improving the user interface for working with the multi-class model for a better understating of the species.

Basic Understanding of project

Through initial understanding and contribution to DeepForest, I have grasped a basic understanding that DeepForest uses Retinanet as a one-stage object detection model that utilizes a focal loss function to address class imbalance during training and which is composed of a backbone network.

Backbone Network

The backbone is responsible for computing a convolutional feature map over an entire input image and is an off-the-self convolutional network to predict individual tree crowns and birds from airborne RGB images. The pre-built model uses a semi-supervised approach to generate millions of moderate-quality annotations using an unsupervised Li DAR tree detection algorithm, followed by manual annotation of RGB images at selected locations.

Depending on what one wants to predict on their data, one requires an adequate backbone network and a certain amount of prediction heads. For a basic classification network, ResNet is the backbone and the fully connected layer as the sole prediction head.

The backbone network convolution and pooling is responsible for extracting a feature map from the image that contains higher-level summarized information. Each head uses this feature map as input to predict its desired outcome. The loss that one optimizes for during training is usually a weighted sum of the individual losses for each prediction head.

Design of DeepForest

Although the main design of this package is for tree detection of a single class which tells the user whether the tree is alive or dead. It uses Retinanet for the creation of a model and Pytorch lightning for the training of model, We can extend the core functionality, and tags with multiple classes according to our need. When creating a DeepForest model object, pass in the expected number of classes and a dictionary of classes that map each numeric class to a character label, moreover model is very sensitive to input resolution of raster images and tends to have low frequency when the resolution of image increases, so there is a need for robust training of model on various resolutions or preprocessing the data and trying to find the appropriate patch size for increasing the model accuracy.