DeepTreeAttention

Tree Species Prediction for the National Ecological Observatory Network (NEON)

Implementation of Hang et al. 2020 Hyperspectral Image Classification with Attention Aided CNNs for tree species prediction.

Model Architecture

Project Organization

├── LICENSE
├── README.md          <- The top-level README for developers using this project.
├── data
│   ├── processed      <- The final, canonical data sets for modeling.
│   └── raw            <- The original, immutable data dump.
│
├── environment.yml   <- Conda requirements
│
├── setup.py           <- makes project pip installable (pip install -e .) so src can be imported
├── src                <- Source code for use in this project.
│   ├── Models         <- Model Architectures

Workflow

There are three main parts to this project, a 1) data module, a 2) model module, and 3) a trainer module. Usually the data_module is created to hold the train and test split and keep track of data generation reproducibility. Then a model architecture is created and pass to the model module along with the data module. Finally the model module is passed to the trainer.

#1) 
data_module = data.TreeData(csv_file="data/raw/neon_vst_data_2021.csv", regenerate=False, client=client)

#2)
model = <create a pytorch NN.module>
m = main.TreeModel(model=model, bands=data_module.config["bands"], classes=data_module.num_classes,label_dict=data_module.species_label_dict)

#3
trainer = Trainer()
trainer.fit(m, datamodule=data_module)

Pytorch Lightning Data Module (data.TreeData)

This repo contains a pytorch lightning data module for reproducibility. The goal of the project is to make it easy to share with others within our research group, but we welcome contributions from outside the community. While all data is public, it is VERY large (>20TB) and cannot be easily shared. If you want to reproduce this work, you will need to download the majority of NEON's camera, HSI and CHM data and change the paths in the config file. For the 'raw' NEON tree stem data see data/raw/neon_vst_2021.csv. The data module starts from this state, which are x,y locations for each tree. It then performs the following actions as an end-to-end workflow.

1. Filters the data to represent trees over 3m with sufficient number of training samples
2. Extract the LiDAR derived canopy height and compares it to the field measured height. Trees that are below the canopy are excluded based on the min_CHM_diff parameter in the config.
3. Splits the training and test x,y data such that field plots are either in training or test.
4. For each x,y stem location the crown is predicted by the tree detection algorithm (DeepForest - https://deepforest.readthedocs.io/).
5. Crops of each tree crown are created and divided into pixel windows for pixel-level prediction.

This workflow does not need to be run on every experiment. If you are satisifed with the current train/test split and data generation process, set regenerate=False

data_module = data.TreeData(csv_file="data/raw/neon_vst_data_2021.csv", regenerate=False)
data_module.setup()

Pytorch Lightning Training Module (data.TreeModel)

Training is handled by the TreeModel class which loads a model from the models folder, reads the config file and runs the training. The evaluation metrics and images are computed and put of the comet dashboard

m = main.TreeModel(model=Hang2020.vanilla_CNN, bands=data_module.config["bands"], classes=data_module.num_classes,label_dict=data_module.species_label_dict)

trainer = Trainer(
    gpus=data_module.config["gpus"],
    fast_dev_run=data_module.config["fast_dev_run"],
    max_epochs=data_module.config["epochs"],
    accelerator=data_module.config["accelerator"],
    logger=comet_logger)
   
trainer.fit(m, datamodule=data_module)

Alive/Dead Filtering

As part of the prediction pipeline, RGB crops are scored as either 'Alive', meanining they have leaves during presumed leaf-on season, or 'Dead', meaning they do not have leaves. To finetune the resent50 model, see src/models/dead.py. The classified data for the Alive/Dead crops can be found in data/raw/dead_train and dead/raw/dead_test.

Dev Guide

In general, major changes or improvements should be made on a new git branch. Only core improvements should be made on the main branch. If a change leads to higher scores, please create a pull request. Any pull requests are expected to have pytest unit tests (see tests/) that cover major use cases.

Model Architectures

The TreeModel class takes in a create model function

m = main.TreeModel(model=Hang2020.vanilla_CNN)

Any model can be specified provided it follows the following input and output arguments

class myModel(Module):
    """
    Model description
    """
    def __init__(self, bands, classes):
        super(myModel, self).__init__()
        <define model architecture here>

    def forward(self, x):
        <forward method for computing loss goes here>
        class_scores = F.softmax(x)
        
        return class_scores

Extending the model

To create a model that takes in new inputs, I strongly recommend sub-classing the existing TreeData and TreeModel classes. For an example, see the MetadataModel in models/metadata.py

#Subclass of the training model
class MetadataModel(main.TreeModel):
    """Subclass the core model and update the training loop to take two inputs"""
    def __init__(self, model, sites,classes, label_dict, config):
        super(MetadataModel,self).__init__(model=model,classes=classes,label_dict=label_dict, config=config)  
    
    def training_step(self, batch, batch_idx):
        """Train on a loaded dataset
        """
        #allow for empty data if data augmentation is generated
        inputs, y = batch
        images = inputs["HSI"]
        metadata = inputs["site"]
        y_hat = self.model.forward(images, metadata)
        loss = F.cross_entropy(y_hat, y)    
        
        return loss

Getting Started (UF - collaboration)

This section is meant solely for members of the idtrees group who have access to the data.

1. Fork this repo and install the conda environment.

conda env create -f=environment.yml
conda activate DeepTreeAttention

2. Update the config.yml

Currently, only members of the ewhite group have permissions to the raw NEON data.

For example:

rgb_sensor_pool: /orange/ewhite/NeonData/*/DP3.30010.001/**/Camera/**/*.tif

This is not a problem, just set

regenerate: False

and it will bypass these steps and use the existing train/test split (e.g. data/processed/train.csv)

You will need to set the correct crop directories

crop_dir: /blue/ewhite/b.weinstein/DeepTreeAttention/crops/

To wherever the crops are saved. This is currently

/orange/idtrees-collab/DeepTreeAttention/crops/

I highly recommend making a comet login. Change

#Comet dashboard
comet_workspace: bw4sz

to your usename and add a .comet.config file to authenticate.

3. Submit a job

Submit a SLURM job

sbatch SLURM/experiment.sh

4. Look at the comet repo for results

The metrics tab has the Micro and Macro Accuracy.