Exploring and Detecting Farm Plots in the Kavango-Zambezi Transfrontier Conservation Area in Collaboration with WWF Germany

This project is about monitoring agricultural practices in the Kavango-Zambezi Transfrontier Conservation Area in collaboration with WWF Germany's Space+Science Team. To do so, it is necessary to know where agricultural areas (i. e. farm plots) are and how their locations and properties change over time. Satellite images (e. g. from Sentinel-2) can be used to identifiy cropland, applying different methods such as pixel-wise classification and instance segmentation.

A report on the project's current status can be found here.

There are two datasets to explore. First, there is a dataset collected during a WWF field campaign, consisting of point coordinates that mark farm plots and spanning the period from 2017 to 2020. These point coordinates have been used by WWF as the basis for manually drawing farm plot polygons using Google Earth in 2022. This notebook describes and explores the field campaign data in detail. Second, there is the KAZA land cover dataset by WWF, constituting a land cover map for the entire KAZA complex based on 2020 data and with 18 distinct land cover classes, including cropland. The data can be downloaded here, including a link to the technical report that describes the dataset and its creation in detail.

It is of particular relevance to explore and detect farm plots in the 6 Bengo regions and the ARISE project within KAZA.

Repository structure:

• earth_engine_python_api: notebooks using the Google Earth Engine Python API
• earth_engine_script: scripts to be run in Google Earth Engine directly
• field_campaign_dataset: notebooks using the field campaign dataset
• kaza_land_cover_dataset: notebooks using the KAZA land cover dataset
• visualizations: created graphs, images and maps

Contents

• What does the region look like from space anyway?
• Cropland mapping using field campaign dataset
• Cropland mapping using KAZA land cover dataset
• Requirements
• References and further information

What does the region look like from space anyway?

The following images show Sentinel-2 RGB mean composites at different temporal resolutions for the sample region of Sioma in 2020, created using Google Earth Engine.

Yearly resolution

Quarterly resolution

Monthly resolution

It is evident that the area is subject to significant variations throughout the seasons of a given year. February seems to be rather clouldy. March seems to be somewhat cloudy. July to November seem to be rather dry and barely vegetated. January, April, May, June and December seem to be the images showing the least amount of clouds and the most amount of vegetation. The growing and harvesting seasons need to be considered as well. The planting usually takes place in November/December. The harvest takes place in April/May.

Cropland mapping using field campaign dataset

The used models are trained and applied to the sample region of Sioma, using Google Earth Engine.

The used land cover classes include:

• Green: vegetation
• Blue: water
• Yellow: farm_plots

The data for the vegetation and water classes consist of hand-drawn points while the farm_plots class represents the point coordinates/polygons from the field campaign dataset mentioned above.

Random forest

Using the points dataset

Using the polygons dataset

Using the points over the polygons dataset seems to yield better results. While the classification of vegetation and water seems to show decent results, it is quite obvious that the same doesn't apply to farm plots. Some farm plot pixels are correctly classified but bare soil, settlement areas and roads for example are also classified as farm plots.

Neural net using the points dataset

Using a more complex model, namely a neural net with 3 hidden layers, doesn't seem to yield significantly better classification results. The classified image can be further explored in a split panel in Google Earth Engine using this link to the Code Editor. The issue of distinguishing farm plots, bare soil, settlements and roads remains. Other land cover classes containing bare soil and human-influenced structures like settlements and roads could be helpful (e. g. KAZA land cover dataset). Instance segmentation, considering the shape of objects, or leveraging the fact that a farm plot pixel looks differently throughout the year could also be more promising paths.

Cropland mapping using KAZA land cover dataset

This notebook is used to sample crop and non-crop points from the KAZA land cover dataset.

Random forest

A random forest classifier is applied to each one of the 6 Bengo regions and also to all of them at once to verify the validity of the sampled points and the resulting classification. Four different balanced datasets are evaluated. A Sentinel-2 cloud-masked mean composite from April 2020 is used as the input image. The features used are the bands B2, B3, B4 and B8 and also NDVI.

Sampling mode	Sample size	Region	Train accuracy	Test accuracy
Random	2,000	Binga	0.96	0.82
Random	2,000	Hwange	0.95	0.66
Random	2,000	Mufunta	0.96	0.73
Random	2,000	Mulobesi	0.98	0.94
Random	2,000	Sichifulo	0.94	0.78
Random	2,000	Zambezi	0.95	0.65
Uniform	2,000	Binga	0.93	0.75
Uniform	2,000	Hwange	0.96	0.75
Uniform	2,000	Mufunta	0.94	0.75
Uniform	2,000	Mulobesi	0.92	0.61
Uniform	2,000	Sichifulo	0.97	0.88
Uniform	2,000	Zambezi	0.96	0.71
Random	20,000	Binga	0.95	0.78
Random	20,000	Hwange	0.95	0.78
Random	20,000	Mufunta	0.96	0.81
Random	20,000	Mulobesi	0.97	0.82
Random	20,000	Sichifulo	0.96	0.85
Random	20,000	Zambezi	0.94	0.72
Uniform	20,000	Binga	0.96	0.74
Uniform	20,000	Hwange	0.96	0.8
Uniform	20,000	Mufunta	0.96	0.74
Uniform	20,000	Mulobesi	0.96	0.77
Uniform	20,000	Sichifulo	0.96	0.8
Uniform	20,000	Zambezi	0.95	0.71
Random	2,000	All	0.96	0.68
Uniform	2,000	All	0.95	0.7
Random	20,000	All	0.95	0.74
Uniform	20,000	All	0.96	0.74

Finally, a crop map is created for the 6 Bengo regions, using the dataset containing 20,000 randomly sampled points. The class distribution in the used dataset is also shown as a reference.

Class distribution in used dataset	Class distribution in predicted crop map

OpenMapFlow

NASA Harvest's OpenMapFlow takes into account that pixel values change throughout the year, uses more than just RGB bands and Sentinel-2 data and applies a pre-trained deep learning model, that can be tuned using data from the respective region of interest, resulting in superior predictive performance compared to using data containing a single timestep.

The model used to create the exemplary maps below includes 2,000 randomly sampled points from the KAZA land cover dataset (and data from the GeowikiLandcover 2017 dataset). The maps have been created using this notebook (using data exported via Google Earth Engine and stored in a Google Cloud bucket) They can be further explored in Google Earth Engine using this link to the Code Editor. The classification threshold is currently set to 0.5 but should be optimized to create the best possible crop masks.

Exemplary 2020 crop probability map (Zambia)	Exemplary 2020 crop mask (Zambia)

The model could be applied to all 6 Bengo regions and the ARISE project to create crop maps for 2020, 2021 and 2022 if it is satisfactory with regard to performance and map quality. Maps for small regions can be created locally or in Google Colab but Google Cloud should be used to create large-scale maps.

Requirements

Assuming conda is installed, run the following to run the Python notebooks:

• create environment: conda env create --file=farm_plot_detection.yaml
• activate environment: conda activate farm_plot_detection

The Google Earth Engine Python API notebooks and the notebook to create maps for small regions using OpenMapFlow are best run directly in Google Colab.

References and further information

Google Earth Engine

• https://developers.google.com/earth-engine/datasets/catalog/COPERNICUS_S2_SR
• https://developers.google.com/earth-engine/guides/tf_examples
• https://developers.google.com/earth-engine/apidocs/ee-algorithms-image-segmentation-snic
• https://www.youtube.com/watch?v=2R0aTaMtYTY

GitHub

• https://github.com/artefactory/medium_satellite_imagery
• https://github.com/chrieke/InstanceSegmentation_Sentinel2
• https://github.com/sentinel-hub/field-delineation
• https://github.com/robmarkcole/satellite-image-deep-learning#segmentation---vegetation-crops--crop-boundaries
• https://github.com/nasaharvest/togo-crop-mask
• https://github.com/nasaharvest/crop-mask
• https://github.com/nasaharvest/openmapflow

Blogs

• https://medium.com/artefact-engineering-and-data-science/leveraging-satellite-imagery-for-machine-learning-computer-vision-applications-d22143f72d94
• https://medium.com/artefact-engineering-and-data-science/applying-machine-learning-algorithms-to-satellite-imagery-for-agriculture-applications-d505eb8df1b1
• https://towardsdatascience.com/farm-segmentation-from-satellite-images-using-holistically-nested-edge-detection-63454a24b164
• https://medium.com/radiant-earth-insights/detecting-agricultural-croplands-from-sentinel-2-satellite-imagery-a025735d3bd8
• https://soilmate.medium.com/%D1%81r%D0%BE%D1%80-field-boundary-detection-approaches-overview-and-main-challenges-53736725cb06
• https://medium.com/onesoil/60-million-fields-and-27-crops-how-we-made-the-onesoil-map-61fe5f3bbb00

Other

• https://sentinelsat.readthedocs.io/en/stable/api_overview.html#lta-products
• https://scihub.copernicus.eu/news/News00592
• https://scihub.copernicus.eu/news/News00382
• https://scihub.copernicus.eu/news/News00596
• https://scihub.copernicus.eu/userguide/