Senegal LCLUC TensorFlow

Python library to process and classify remote sensing imagery by means of GPUs and CPU parallelization for high performance and commodity base environments. This repository focuses in using CNNs for the inference of very high-resolution remote sensing imagery in Senegal.

Figure 1. Study area of this repository.

We are currently working on tutorials and documentations. Feel to follow this repository for documentation updates and upcoming tutorials.

Science Questions

• Can 2m VHR imagery enhance our understanding of changes in the extent, intensity and land use of agriculture and forestry in Senegal?
• How can we better take advantage of NASA’s HEC resources to apply Deep learning for land cover change monitoring?
• Can we scale-up Unet CNNs to compensate for the diversity of landscapes and images to map regional land cover?

Figure 2. Wet and dry seasonality for cluster of typical fields in Senegal, Photos from Collaborator Gray Tappan

Downloading the Container

The container for this work can be downloaded from DockerHub. The container is deployed on a weekly basis to take care of potential OS vulnerabilities. All CPU and GPU dependencies are baked into the container image for end-to-end processing.

singularity build --sandbox /lscratch/$USER/container/tensorflow-caney docker://nasanccs/tensorflow-caney:latest

Data Preprocessing

The data used in this work is provided by NGA through the NextView agreement. As long as you are part of this agreement, you can get access to the WorldView imagery we have processed. We also work with Planet imagery as part of this project, which is free of access through the NICFI program.

Generate Tappan Squares

To generate training and small samples of data we generated the so called Tappan squares in honor of our colleague Gray Tappan. These Tappan squares are 5000x5000 pixel squares, for an area of 10000x10000 m^2. Generating Tappan squares can be achieved with the tapann_pipeline_cli.py script. Its execution is as follows:

singularity exec --env PYTHONPATH="/explore/nobackup/people/$USER/development/senegal-lcluc-tensorflow:/explore/nobackup/people/$USER/development/tensorflow-caney:/explore/nobackup/people/jacaraba/development" --nv -B /explore/nobackup/projects/ilab,/explore/nobackup/projects/3sl,$NOBACKUP,/explore/nobackup/people /lscratch/$USER/container/tensorflow-caney python /explore/nobackup/people/$USER/development/senegal-lcluc-tensorflow/senegal_lcluc_tensorflow/view/tappan_pipeline_cli.py -c /explore/nobackup/people/$USER/development/senegal-lcluc-tensorflow/projects/tappan_generation/configs-srlite/tappan_06.yaml

Land Cover CNN Workflow

In this workflow we perform land cover segmentation using CNNs. The main steps from this workflow are the preprocess, train, predict, and validate steps. The CLI binary is used to manage the execution of these steps in the pipeline.

Full Pipeline

Here we run the full pipeline for preprocess, train, and predict.

singularity exec --env PYTHONPATH="/explore/nobackup/people/$USER/development/senegal-lcluc-tensorflow:/explore/nobackup/people/$USER/development/tensorflow-caney" --nv -B /explore/nobackup/projects/ilab,/explore/nobackup/projects/3sl,$NOBACKUP,/lscratch,/explore/nobackup/people /lscratch/$USER/container/tensorflow-caney python /explore/nobackup/people/$USER/development/senegal-lcluc-tensorflow/senegal_lcluc_tensorflow/view/landcover_cnn_pipeline_cli.py -c /explore/nobackup/people/$USER/development/senegal-lcluc-tensorflow/projects/land_cover/configs/experiments/2023-AccuracyIncrease/global_standardization_256_crop_4band_short.yaml -d /explore/nobackup/people/$USER/development/senegal-lcluc-tensorflow/projects/land_cover/configs/experiments/2023-AccuracyIncrease/land_cover_512_otcb_50TS_cas-wcas-short.csv -s preprocess train predict

Another example:

singularity exec --env PYTHONPATH="/explore/nobackup/people/$USER/development/senegal-lcluc-tensorflow:/explore/nobackup/people/$USER/development/tensorflow-caney" --nv -B /explore/nobackup/projects/ilab,/explore/nobackup/projects/3sl,$NOBACKUP,/lscratch,/explore/nobackup/people /lscratch/$USER/container/tensorflow-caney python /explore/nobackup/people/$USER/development/senegal-lcluc-tensorflow/senegal_lcluc_tensorflow/view/landcover_cnn_pipeline_cli.py -c /explore/nobackup/people/$USER/development/senegal-lcluc-tensorflow/projects/land_cover/configs/experiments/2023-AccuracyIncrease/8bit_scale_256_crop_4band_short.yaml -d /explore/nobackup/people/$USER/development/senegal-lcluc-tensorflow/projects/land_cover/configs/experiments/2023-AccuracyIncrease/land_cover_512_otcb_50TS_cas-wcas-8bit-short.csv -s preprocess train predict

Test

Generate metrics and statistics using test data.

singularity exec --env PYTHONPATH="/explore/nobackup/people/$USER/development/senegal-lcluc-tensorflow:/explore/nobackup/people/$USER/development/tensorflow-caney" --nv -B /explore/nobackup/projects/ilab,/explore/nobackup/projects/3sl,$NOBACKUP,/explore/nobackup/people /explore/nobackup/projects/ilab/containers/tensorflow-caney-2023.05 python /explore/nobackup/people/$USER/development/senegal-lcluc-tensorflow/senegal_lcluc_tensorflow/view/landcover_cnn_pipeline_cli.py -c /explore/nobackup/people/$USER/development/senegal-lcluc-tensorflow/projects/land_cover/configs/experiments/2023-GMU-V2/batch1/eCAS-wCAS-otcb-40/eCAS-wCAS-otcb-40.yaml -t '/explore/nobackup/projects/3sl/labels/landcover/2m_all_fixed/*.tif' -s test

Validate

Generate metrics and statistics using validation data.

singularity exec --env PYTHONPATH="/explore/nobackup/people/jacaraba/development/senegal-lcluc-tensorflow:/explore/nobackup/people/jacaraba/development/tensorflow-caney",PROJ_LIB='/usr/share/proj' --nv -B /explore/nobackup/projects/ilab,/explore/nobackup/projects/3sl,$NOBACKUP,/lscratch,/explore/nobackup/people /explore/nobackup/projects/ilab/containers/above-shrubs.2023.07 python /explore/nobackup/people/jacaraba/development/senegal-lcluc-tensorflow/senegal_lcluc_tensorflow/view/landcover_cnn_pipeline_cli.py -v '/explore/nobackup/projects/3sl/data/Validation/3sl-validation-database-2023-10-05-all-row-based_all-agreed.gpkg' -c /explore/nobackup/people/jacaraba/development/senegal-lcluc-tensorflow/projects/land_cover/configs/experiments/2023-GMU-V2/batch2/eCAS-wCAS-ETZ-otcb-80_repeat3/eCAS-wCAS-ETZ-otcb-80_repeat3.yaml -s validate

In some cases we need to perform larger scale validations (for example, N number of subdirectories). The following script allows you to perform the validation using the subdirectory where the configuration file is available. All we ned to do is to point to the main directory that houses all other configuration subdirectories.

This is an example for general TOA imagery using the GMU experiments for the paper:

bash validation_gmu.sh \
    "/explore/nobackup/people/$USER/development/senegal-lcluc-tensorflow/projects/land_cover/configs/experiments/2023-GMU-V2/batch1/*/*.yaml" \
    "/explore/nobackup/projects/3sl/data/Validation/3sl-validation-database-2023-10-05-all-row-based_all-agreed.gpkg"

An an example for the Surface Reflectance experiments:

bash validation_srlite.sh \
    "/explore/nobackup/people/$USER/development/senegal-lcluc-tensorflow/projects/land_cover/configs/experiments/NASA/2023-surface_reflectance/ard_srlite_toa/*/*.yaml" \
    "/explore/nobackup/projects/3sl/data/Validation/3sl-validation-database-2023-10-05-all-row-based_all-agreed.gpkg"

Compositing

After we generate predictions for the entire study area, we need to proceed to create composites. Below you will find the documentation to perform the compositing steps. This pipeline has 3 main steps:

1. Build footprints
2. Extract metadata
3. Build composite

Below you will find examples on how to run each one of these.

Build footprints

singularity exec --env PYTHONPATH="/explore/nobackup/people/jacaraba/development/vhr-composite:/explore/nobackup/people/jacaraba/development/senegal-lcluc-tensorflow" --nv -B $NOBACKUP,/lscratch,/explore/nobackup/people,/explore/nobackup/projects /explore/nobackup/projects/ilab/containers/tensorflow-caney-2023.05 python /explore/nobackup/people/jacaraba/development/senegal-lcluc-tensorflow/senegal_lcluc_tensorflow/view/landcover_composite_pipeline_cli.py -c /explore/nobackup/people/jacaraba/development/senegal-lcluc-tensorflow/projects/composite/configs/composite_cas.yaml -t /explore/nobackup/people/jacaraba/development/senegal-lcluc-tensorflow/projects/composite/configs/tile_lists/test_tile_0.txt -s build_footprints

Extract metadata

singularity exec --env PYTHONPATH="/explore/nobackup/people/jacaraba/development/vhr-composite:/explore/nobackup/people/jacaraba/development/senegal-lcluc-tensorflow" --nv -B $NOBACKUP,/lscratch,/explore/nobackup/people,/explore/nobackup/projects /explore/nobackup/projects/ilab/containers/tensorflow-caney-2023.05 python /explore/nobackup/people/jacaraba/development/senegal-lcluc-tensorflow/senegal_lcluc_tensorflow/view/landcover_composite_pipeline_cli.py -c /explore/nobackup/people/jacaraba/development/senegal-lcluc-tensorflow/projects/composite/configs/composite_cas.yaml -t /explore/nobackup/people/jacaraba/development/senegal-lcluc-tensorflow/projects/composite/configs/tile_lists/test_tile_0.txt -s extract_metadata

Build composite

singularity exec --env PYTHONPATH="/explore/nobackup/people/jacaraba/development/vhr-composite:/explore/nobackup/people/jacaraba/development/senegal-lcluc-tensorflow" --nv -B $NOBACKUP,/lscratch,/explore/nobackup/people,/explore/nobackup/projects /explore/nobackup/projects/ilab/containers/tensorflow-caney-2023.05 python /explore/nobackup/people/jacaraba/development/senegal-lcluc-tensorflow/senegal_lcluc_tensorflow/view/landcover_composite_pipeline_cli.py -c /explore/nobackup/people/jacaraba/development/senegal-lcluc-tensorflow/projects/composite/configs/composite_cas.yaml -t /explore/nobackup/people/jacaraba/development/senegal-lcluc-tensorflow/projects/composite/configs/tile_lists/tile_0.txt -s composite

Land Use 1D CNN Workflow

In this workflow we perform land use object segmentation using CNNs.

Setup

singularity exec --env PYTHONPATH="/explore/nobackup/people/$USER/development/senegal-lcluc-tensorflow:/explore/nobackup/people/$USER/development/tensorflow-caney" --nv -B /explore/nobackup/projects/ilab,/explore/nobackup/projects/3sl,$NOBACKUP,/lscratch,/explore/nobackup/people /lscratch/$USER/container/tensorflow-caney python /explore/nobackup/people/$USER/development/senegal-lcluc-tensorflow/senegal_lcluc_tensorflow/view/landuse_cnn_pipeline_cli.py -c /explore/nobackup/people/$USER/development/senegal-lcluc-tensorflow/projects/land_use/configs/landuse.yaml --gee-account 'id-sl-senegal-service-account@ee-3sl-senegal.iam.gserviceaccount.com' --gee-key '/home/$USER/gee/ee-3sl-senegal-8fa70fe1c565.json' -s setup

CLoud Masking Imagery

In this workflow we use the vhr-cloudmask software to generate cloud masks of our WorldView imagery. The command to run cloud mask across the imagery from the PRISM GPU cluster is as follows, where:

• '-o' is the output directory
• '-r' is a list of regex where data files live
• '-s' is the pipeline step to perform which in this case is 'predict'

for i in {0..64}; do sbatch --mem-per-cpu=10240 -G1 -c10 -t05-00:00:00 -J clouds -q ilab --wrap="singularity exec --nv -B $NOBACKUP,/explore/nobackup/people,/explore/nobackup/projects /explore/nobackup/projects/ilab/containers/vhr-cloudmask.sif vhr-cloudmask-cli -o '/explore/nobackup/projects/3sl/products/cloudmask/v2' -r '/explore/nobackup/projects/3sl/data/Tappan/new_dates/Tappan*.data.tif' '/explore/nobackup/projects/3sl/data/VHR/new_evhr/5-toas/*toa.tif' -s predict"; done

TOA vs ARD vs SRLite

The following documentation pertains to the TOA vs ARD vs SRLite experiments. The process requires several pieces and parts, but here is a summary of the process for reproducibility. The overall process looks like this:

1. Identify TOA scenes to process.
2. Run EVHR.
3. Run SRLite.
4. Generate Tappan squares.
5. Run the different CNN experiments (training, inference, validation).
6. Visualize and summarize.

Below you will find the different commands to run each step from the pipeline.

Identify TOA scenes to process

In this step we generate a CSV file with the respective TOA scenes available in ADAPT/Explore. For this we use a notebook provided by mwooten that queries the NGA database.

Run EVHR

For this step, we need the list of NTF files together with the output directory where we want to store the output data. The output directory is to /explore/nobackup/projects/3sl/data/EVHR/<study_area>, where study area is one of CAS, ETZ, SRV.

ILAB Nodes:

cd /explore/nobackup/people/jacaraba/development/senegal-lcluc-tensorflow/scripts/evhr
./run_evhr_exec /explore/nobackup/people/jacaraba/development/senegal-lcluc-tensorflow/scripts/evhr/input_files_all_senegal/CAS_M1BS_scenes/subset_2010/CAS_M1BS_scenes_noDups_2010_1.csv /explore/nobackup/projects/3sl/data/EVHR/CAS /explore/nobackup/people/iluser/ilab_containers/evhr_4.0.0.sif

PRISM Nodes:

salloc --ntasks 1 --cpus-per-task 40 -G1 -t 05-00:00:00 -J lightning
cd /explore/nobackup/people/jacaraba/development/senegal-lcluc-tensorflow/scripts/evhr
./run_evhr_exec /explore/nobackup/people/jacaraba/development/senegal-lcluc-tensorflow/scripts/evhr/input_files_all_senegal/CAS_M1BS_scenes/subset_2010/CAS_M1BS_scenes_noDups_2010_1.csv /explore/nobackup/projects/3sl/data/EVHR/CAS /lscratch/jacaraba/container/evhr-container

The following are the scenes we need to process via EVHR.

/explore/nobackup/people/jacaraba/development/senegal-lcluc-tensorflow/scripts/evhr/input_files_all_senegal/CAS_M1BS_scenes/CAS_M1BS_scenes_noDups_2010.csv
/explore/nobackup/people/jacaraba/development/senegal-lcluc-tensorflow/scripts/evhr/input_files_all_senegal/CAS_M1BS_scenes/CAS_M1BS_scenes_noDups_2011.csv
/explore/nobackup/people/jacaraba/development/senegal-lcluc-tensorflow/scripts/evhr/input_files_all_senegal/CAS_M1BS_scenes/CAS_M1BS_scenes_noDups_2012.csv
/explore/nobackup/people/jacaraba/development/senegal-lcluc-tensorflow/scripts/evhr/input_files_all_senegal/CAS_M1BS_scenes/CAS_M1BS_scenes_noDups_2013.csv
/explore/nobackup/people/jacaraba/development/senegal-lcluc-tensorflow/scripts/evhr/input_files_all_senegal/CAS_M1BS_scenes/CAS_M1BS_scenes_noDups_2014.csv
/explore/nobackup/people/jacaraba/development/senegal-lcluc-tensorflow/scripts/evhr/input_files_all_senegal/CAS_M1BS_scenes/CAS_M1BS_scenes_noDups_2015.csv
/explore/nobackup/people/jacaraba/development/senegal-lcluc-tensorflow/scripts/evhr/input_files_all_senegal/CAS_M1BS_scenes/CAS_M1BS_scenes_noDups_2016.csv
/explore/nobackup/people/jacaraba/development/senegal-lcluc-tensorflow/scripts/evhr/input_files_all_senegal/CAS_M1BS_scenes/CAS_M1BS_scenes_noDups_2017.csv
/explore/nobackup/people/jacaraba/development/senegal-lcluc-tensorflow/scripts/evhr/input_files_all_senegal/CAS_M1BS_scenes/CAS_M1BS_scenes_noDups_2018.csv
/explore/nobackup/people/jacaraba/development/senegal-lcluc-tensorflow/scripts/evhr/input_files_all_senegal/CAS_M1BS_scenes/CAS_M1BS_scenes_noDups_2019.csv
/explore/nobackup/people/jacaraba/development/senegal-lcluc-tensorflow/scripts/evhr/input_files_all_senegal/CAS_M1BS_scenes/CAS_M1BS_scenes_noDups_2020.csv
/explore/nobackup/people/jacaraba/development/senegal-lcluc-tensorflow/scripts/evhr/input_files_all_senegal/CAS_M1BS_scenes/CAS_M1BS_scenes_noDups_2021.csv
/explore/nobackup/people/jacaraba/development/senegal-lcluc-tensorflow/scripts/evhr/input_files_all_senegal/CAS_M1BS_scenes/CAS_M1BS_scenes_noDups_2022.csv
/explore/nobackup/people/jacaraba/development/senegal-lcluc-tensorflow/scripts/evhr/input_files_all_senegal/CAS_M1BS_scenes/CAS_M1BS_scenes_noDups.csv

Authors

• Jordan Alexis Caraballo-Vega, jordan.a.caraballo-vega@nasa.gov
• Margaret Wooten, margaret.wooten@nasa.gov
• Caleb S. Spradlin, caleb.s.spradlin@nasa.gov
• Minh Tri Le, mle35@gmu.edu

Installation

See the build guide.

Contributing

Please see our guide for contributing to terragpu.

References

[1] Raschka, S., Patterson, J., & Nolet, C. (2020). Machine learning in python: Main developments and technology trends in data science, machine learning, and artificial intelligence. Information, 11(4), 193.

[2] Paszke, Adam; Gross, Sam; Chintala, Soumith; Chanan, Gregory; et all, PyTorch, (2016), GitHub repository, https://github.com/pytorch/pytorch. Accessed 13 February 2020.

[3] Caraballo-Vega, J., Carroll, M., Li, J., & Duffy, D. (2021, December). Towards Scalable & GPU Accelerated Earth Science Imagery Processing: An AI/ML Case Study. In AGU Fall Meeting 2021. AGU.