BA-Net: A deep learning approach for mapping and dating burned areas using temporal sequences of satellite images

Over the past decades, methods for burned areas mapping and dating from remote sensing imagery have been the object of extensive research. The limitations of current methods, together with the heavy pre-processing of input data they require, make them difficult to improve or apply to different satellite sensors. Here, we explore a deep learning approach based on daily sequences of multi-spectral images, as a promising and flexible technique that can be applicable to observations with various spatial and spectral resolutions. We test the proposed model for five regions around the globe using input data from VIIRS 750 m bands resampled to a 0.01º spatial resolution grid. The derived burned areas are validated against higher resolution reference maps and compared with the MCD64A1 Collection 6 and FireCCI51 global burned area datasets. We show that the proposed methodology achieves competitive results in the task of burned areas mapping, despite using lower spatial resolution observations than the two global datasets. Furthermore, we improve the task of burned areas dating for the considered regions of study when compared with state-of-the-art products. We also show that our model can be used to map burned areas for low burned fraction levels and that it can operate in near-real-time, converging to the final solution in only a few days. The obtained results are a strong indication of the advantage of deep learning approaches for the problem of mapping and dating of burned areas and provide several routes for future research.

Install

libhdf4-dev is required for pyhdf to read .hdf4 files.

sudo apt install -y libhdf4-dev

pip install banet

Dataset

The dataset used to train the model and the pretrained weights are available at https://drive.google.com/drive/folders/1142CCdtyekXHc60gtIgmIYzHdv8lMHqN?usp=sharing. Notice that the size of the full dataset is about 160 GB. You can, however, donwload individual regions in case you want to test with a smaller dataset.

Generate predictions for a new region in 5 steps

The procedure to generate predictions for a new region or for a different period of the existing regions is straightforward.

• Step 1. Define a .json file with region name, bounding boxes and spatial resolution. Region definition files are by default on data/regions and should be named as R_{name}.json, where name is the name you give to the region inside the file. For example for Iberian Peninsula region the file data/regions/R_PI.json contains the following: {"name": "PI", "bbox": [-10, 36, 5, 44], "pixel_size": 0.01}.

• Step 2. To download the reflectance data the command line script banet_viirs750_download can be used. However, in order to use it, you need to first register at the website (https://ladsweb.modaps.eosdis.nasa.gov/) and generate an authentication token.

• Step 3. Next you need to download the VIIRS active fire data for the region you seletected. This procedure is manual but you should be able to request data for the entire temporal window in one go. To do that go to https://firms.modaps.eosdis.nasa.gov/download/, select Create New Request, select the region based on your region bounding box, for fire data source select VIIRS, then select the date range and finally .csv for the file format. You should receive an email with the confirmation and later another with the link to download the data. If not go back to the download page enter the email you used for the request and choose Check Request Status. If it is completed the download link will appear. Once you have the file place it in data/hotspots and name it hotspots{name}.csv where name is the name of the region as in the .json file defined in Step 1.

• Step 4. Now that you have all the data, you can use the command line script banet_create_dataset.

• Step 5. Finally you can use the banet_predict_monthly command line script to generate the model outputs.

Note: Some examples of usage for the command line tools are available in the documentation.

Train the model from scratch

To train the model you need a dataset of image tiles and the respective targets. The data for the 5 study regions is available for download at https://drive.google.com/drive/folders/1142CCdtyekXHc60gtIgmIYzHdv8lMHqN?usp=sharing. In case you want to train a model on different regions the procedure to collect VIIRS data is described on Generate predictions for a new region in 5 steps above.

• Step 1. In case you opted for new regions you need to collect MCD64A1 collection 6 burned areas to use as targets. MCD64A1 collection 6 data can be downloaded from ftp://ba1.geog.umd.edu/ server (description and credentials available on http://modis-fire.umd.edu/files/MODIS_C6_BA_User_Guide_1.2.pdf section 4). Once you log into the server go to Collection 6/TIFF folder and download data for the window or windows covering your region (Figure 2 on the user guide shows the delineation of the windows).

• Step 2. Once you have the data you can generate a dataset with banet_create_dataset and banet_dataset2tiles command line tools. If you downloaded the dataset already in .mat files format provided in the url above then you just use the banet_dataset2tiles.

• Step 3. To train the model the banet_train_model command line tool is provided. To use sequences of 64 days with 128x128 size tiles and batch size of 1, you need a 8 GB GPU. You can try reducing the sequence length if your GPU has less memory or increase the batch size otherwise.

Note: Some examples of usage for the command line tools are available in the documentation.

Fine-tune the model for a specific region or for other data source (transfer learning)

It is possible to fine-tune the trained model weights to a specific region or using a different source for the input data (e.g., data from VIIRS 375m bands or data from another satellite). The easiest way to include a new dataset is to write a Dataset class similar to Viirs750Dataset or the other dataset classes on banet.data. Once you have the new dataset you can follow the Train the model from scratch guideline and the only change you need to make is to make sure you load the pretrained weights before starting to train.

Troubleshooting

If you find any bug or have any question regarding the code or applications of BA-Net you can navigate to Issues tab and create a new Issue describing your problem. I will try to answer as soon as possible.

Citation

@article{pinto2020banet,
  title={A deep learning approach for mapping and dating burned areas using temporal sequences of satellite images},
  author={Pinto, Miguel M and Libonati, Renata and Trigo, Ricardo M and Trigo, Isabel F and DaCamara, Carlos C},
  journal={ISPRS Journal of Photogrammetry and Remote Sensing},
  volume={160},
  pages={260--274},
  year={2020},
  publisher={Elsevier}
}