Sentinel-2 Active Fire Segmentation: Analyzing Convolutional and Transformer Architectures Knowledge Transfer, Fine-Tuning and Seam-Lines
This work analyze three different architectures (U-net, Deeplab v3+ and SegFormer) in regard of transfer learning for segment active fire using remote sensing images. In this study we pre-trained the networks using Landsat-8 images and then perform a transfer learning to Sentinel-2 using manually annotated images. To pre-train the base networks we use the dataset built by Pereira et al. (2021), available on GitHub/Google Drive. To fine-tune we use the dataset built by Fusioka et al. (2024) available on GitHub/Google Drive. We evaluated the performance of the networks for active fire segmentation when performing the transfer learning and also the performance when facing images with seam-lines that causes many false positives detections. Experiments show that the proposed method achieves F1-scores of up to 88.4% for Sentinel-2 images, outperforming three threshold-based algorithms by at least 19%.
Gabriel Henrique de Almeida Pereira
To pre-train the networks we use the Landsat-8 of the dataset built by Pereira et al. (2021). The dataset consists of images from around the world with active fire and their respective masks produced by three threshold algorithms and their combination by intersection and majority voting. In this work we use only the majority voting masks.
With the networks trained on Landsat-8 images, we fine-tune then using Sentinel-2 images manually annotated, we use the dataset built by Fusioka et al. (2024). This dataset consist of 26 Sentinel-2 images manually annotated, which has 22 images with at least one fire pixel and 4 images without any fire pixel. While the dataset offer the images in patches (images cropped in 256x256 pixels), we choose to discard the patches with only no-data (black images) or partially with no-data.
In addition to these dataset we analyzed a set of Sentinel-2 images without fire, but with seam-lines (artifacts caused by the composition of multiple captures, a known issue that results in misaligned bands visible in regions with clouds) that cause false positives in the active fire segmentation task. We made this image available on Google Drive. The link contain a zip file with patches with 256x256 pixels containing the seam-lines. In addition to the patches, we also provide in the zip file masks for the patches generated by tree active fire threshold-based algorithms, namely Kato-Nakamura, Liu and Murphy. In order to save space, we didn't provide the manual annotations for active fire for those patches, since in our visual inspections no fire pixel was identified, therefore all masks are black images.
We have made a small change in the dataset provided by Fusioka et al. (2024), we chose to not use the patches with "no-data". We provide two distinct scripts to adjust the dataset, the first script is the src/utils/crop_images.py
this script can be used to crop the Sentinel-2 scenes available in the dataset. We also provide the script src/utils/remove_no_data_patches.py
that will remove any patch with "no-data" from the samples.
We trained a U-net, Deeplab v3+ and a SegFormer using the Landsat-8 images. To train the base networks we provide and script src/train_landsat.py
that can be used to train the base networks. This script use the configuration defined in the src/landsat_config.py
to train the network defined in the constant MODEL
. You can override this configuration using the argument --model
when invoking the script. You can also change the batch size (--batch-size
), learning rate (--lr
), number of epochs (--epochs
) and early stopping patience (--early-stopping-patience
). The final models will be saved on the folder defined in the constant LANDSAT_OUTPUT_DIR
in the configuration script. You need to set the constant IMAGES_PATH
to the path with the Landsat-8 patches and the constant MASKS_PATH
to the path with the masks.
A example of invoking the training script to train the U-net:
python train_landsat.py --model unet
For training the DeepLabV3+:
python train_landsat.py --model deeplabv3+
For training the SegFormer (B0):
python train_landsat.py --model SegFormerB0
The implementation of these models can be found online: U-net, DeepLab v3+ , SegFormer.
After training the base models you can fine-tune them using the Sentinel-2 images. First, you need to define the CSV files that contains the folds that will be used to train, validate and test the models. Use the script src/generate_folds.py
to create the this files. You can change the configuration of this script in the file src/config.py
. The constant MANUAL_ANNOTATIONS_IMAGE_PATH
must point to the folder holding the Sentinel-2 patches, the constant MANUAL_ANNOTATIONS_MASK_PATH
must point to the manual annotations and the SEAMLINE_IMAGES_PATH
must point to the folder with the seam-line patches. The code will check your images, counting the fire pixels in each image, a summary will be saved in the file defined in the CSV_NUM_FIRE_PIXELS_PER_PATCH_PATH
constant. The number of folds is defined by the constant NUM_FOLDS
the default value is 5. If you want to use stratified folds, based in the categories fire
, seam-line
and no-fire
you can set the constant STRATIFIED_FOLDS
to True
, the default is False
. After setting the configurations in the config.py
you can run the script:
python generate_folds.py
This code will set one split to test and one split to validation, all others splits will be used to train. If you don't want to define a validation fold you can set GENERATE_VALIDATION_FOLD
to False
.
With the folds defined you can fine-tune the network using the Sentinel-2 images. The script src/transfer_learning.py
can be used to this task. The configurations used in this script is also defined in the src/config.py
script. Alternatively, you can change the default configuration using the arguments available in the fine-tuning script. You can use the --model
argument to define the base model (unet, deeplabv3+ or SegFormerB0) to fine-tune. The --csv-folds-dir
argument can be used to point to the folder with the csv files with the folds definition, if you want to use specify the folds to be used you can set them in the --fold
argument. You can also change the number of epochs to fine-tune the model with the argument --epochs
, if you want to use the networks without any fine-tuning you can set the number of epochs to zero, this will only evaluate the model. Alternatively you can pass the argument --no-tuning
, this will disable the fine-tuning step and execute only the evaluation.
When running this script it will fine-tuning and evaluate the model using the defined folds. After the fine-tuning step it will be saved the history in a json file inside the folder defined in the OUTPUT_RESULTS_TRANSFER_LEARNING_PATH
constant, the weights will be saved in the folder defined in OUTPUT_WEIGHTS_TRANSFER_LEARNING_PATH
, the results of the evaluation over the test fold will be saved in the folder OUTPUT_RESULTS_TRANSFER_LEARNING_PATH
as a json file.
We evaluated tree thresholding algorithms in the task of active fire segmentation, the algorithms was proposed by Kato and Nakamura, Liu et. al and Murphy et. al. The masks for each of these methods are available in the dataset. In order to evaluate the performance of these methods for active fire segmentation you can use the script src/evaluate_thresholding_algorithms.py
, this script also rely on the file src/config.py
for the configurations. The results will be save in a csv file in the path define in the constant OUTPUT_RESULTS_METHODS_PREDICTIONS_PATH
. To run this script you can run:
python evaluate_thresholding_algorithms.py
Full article is available in IEEEXplore. Bibtex citation:
@ARTICLE{Fusioka2024,
author={Fusioka, Andre M. and Pereira, Gabriel H. D. A. and Nassu, Bogdan T. and Minetto, Rodrigo},
journal={IEEE Geoscience and Remote Sensing Letters},
title={Sentinel-2 Active Fire Segmentation: Analyzing Convolutional and Transformer Architectures, Knowledge Transfer, Fine-Tuning and Seam-Lines},
year={2024},
volume={},
number={},
pages={1-1},
keywords={Remote sensing;Earth;Artificial satellites;Image segmentation;Satellites;Transfer learning;Training;active fire segmentation;transfer learning;fine-tuning;Sentinel-2 imagery;seam-lines},
doi={10.1109/LGRS.2024.3443775}
}