DS3Lab / forest-prediction

🛰 Deep learning for deforestation classification and forecasting in satellite imagery

forest-prediction

🛰Deep learning for deforestation classification and forecasting in satellite imagery

Overview

In this repository we provide implementations for:

1. Data scraping (Tile services and Google Earth Engine)
2. Forest prediction (Semantic Segmentation)
3. Video prediction (Lee et al, 2018)
4. Image to image translation (Isola et al, 2017)

Installation

$ git clone https://github.com/DS3Lab/forest-prediction.git
$ cd forest-prediction/semantic_segmentation/unet
$ conda create --name forest-env python=3.7
$ ./install.sh
$ source activate forest-env

Running

You can train the models for semantic segmentation by simply running:

(forest-env) $ cd semantic_segmentation/unet
(forest-env) $ python train.py -c {config_path} -d {gpu_id}

For multi-GPU training, set gpu_id to a comma-separated list of devices, e.g. -d 0,1,2,3,4 This will produce a file having the time in which the script was executed as the folder name. It will be saved in the "save_dir" value from the JSON file, under "trainer". Under save_dir, it will create a log file, where you can check Tensorboard, and a model file, where the model is going to be stored.

Testing

You can test the models for semantic segmentation by running:

(forest-env) $ python simple_test.py -r {model_saved_path/model.pth} -d {gpu_id}

It will run the predictions and save the corresponding outputs in model_saved_path. To keep an order of the images, set both batch_size and num_workers to 1.

Configuration

You can change the type of model used, and its configuration by altering (or creating) a config.json file.

Structure of config.json

The fields of the config file are self explanatory. We explain the most important ones.

• name: indicates the name of the experiment. It is the folder in which both the training logs and models are going to be stored
• n_gpu: for multi-GPU training, it is necessary to specify how many gpus it is going to use. For instance, if the user specifies -d 0,1, in order to use both gpus n_gpu needs to be set up to 2. If it is set up to 1, it will only use gpu 0, if it is set up to a number higher than 2, then it will yield an error.
• arch: it specifies the model that will be used for training/testing purposes.
• data_loader_train and data_loader_val: data loaders for training and validation purposes. For testing, only data_loader_val is used.