Omdena successfully combines AI and ML methodologies with open source, low resolution satellite imagery to create actionable solutions for powerful insights.
Every day, millions of images are captured from space by an ever-growing number of satellites. Satellite imagery enables regular remote monitoring of our ecosystem like never before, providing essential information on the current situation on the ground, as well as on environmental changes, global poverty, urban, agricultural, and economic development, human movement, and other global transitions.
Properly processed and analyzed, Earth observation data can offer invaluable insights to inform policy decisions, solve problems, and help direct humanitarian efforts. The European Space Agency (ESA), for instance, has shared how satellite data can be used to address each of the UN’s 17 Sustainable Development Goals.
Exponential growth in the availability of Earth observation data has been driven by the open-source policy adopted by the ESA, NASA, and the U.S. Geological Survey (USGS) as well as by numerous new fleets of small commercial satellites. The potential of remote sensing through Earth observation is growing rapidly with innovations in data science, cloud computing, computer vision, AI, machine learning, and other methods for processing unstructured satellite data.
However, the expense of high-resolution satellite imagery and the lack of expertise necessary to leverage this data and develop appropriate use cases prevent many organizations from taking advantage of these insights.
In a series of six short-term (2 months) projects conducted over the course of 2019, Omdena demonstrated how combining AI and ML methodologies with open source, low-resolution satellite imagery can generate data-driven insights to guide organizations in decision making, planning, and response. Our partners have benefited from the actionable solutions created by Omdena data scientists.
Omdena’s data scientists work with our partner organizations to craft custom approaches tailored to their specific challenges. We help define use cases, develop tools to process affordable, publicly available low-resolution satellite imagery, and augment this data with ground truth from other sources. Our AI models triangulate multiple data sets for high accuracy predictions, creating deployable and ethical solutions for our partners.
Over the past year, Omdena has worked with leading global organizations such as the UN Refugee Agency (UNHCR), the World Food Program, and Impact Hub, as well as with technology startups and NGOs around the world. Our projects combine publicly available satellite imagery with a variety of AI methodologies to help our partners address issues of global poverty, hunger, disaster response, wildfire prevention, humanitarian relief, and space exploration. We have partnered with organizations working in Somalia, Nigeria, Brazil, Sweden, Turkey, Nepal, and even on Mars to craft use cases and address global challenges by identifying and labeling crops, trees, buildings, conflict zones, areas in need of electricity, and anomalies on other planets.
Our process involves working in collaboration with our partners to define the parameters of the problem. We next identify useful data sources – navigating a variety of satellite inputs, and combining them with ground truth data provided by our partners and other publicly available sources such as street maps, energy grids, and population data. We develop ML models that are able to harvest, process, and categorize unstructured data, and triangulate this information with other data points. Our final products include prototypes, data visualizations, interactive maps, and other tools for immediate use by our partners. Many of the solutions developed by Omdena are reproducible and can be leveraged for similar scenarios in other regions of the world.
The six case studies below provide examples of our work
In collaboration with Renewable Africa 365, Omdena data scientists combined satellite imagery, energy grid analysis, population analysis, and AI to identify where Nigeria’s energy poverty crisis is most dire and where solar power is most likely to be effective.
Working with the World Food Program (WFP) in Nepal, Omdena data scientists combined data from multiple satellite bands to create ML models capable of classifying rice and wheat crop fields with an accuracy approaching 89%. This information will help the WFP improve resource allocation, accelerate the growth of staple foods, and reduce hunger in Nepal. Omdena researchers aggregated learnings from this project into a Guide to Using Satellite Imagery in Agricultural Applications that can be leveraged by data scientists worldwide.
In collaboration with Istanbul’s ImpactHub innovation center, Omdena data scientists identified the problem – emergency response in an earthquake-prone region – and then the solution. Our data scientists combined satellite imagery of Istanbul with street map data in order to build a tool that facilitates family reunification by indicating the shortest and safest route between two points after an earthquake.
Working with the UNHCR in Somalia, Omdena data scientists effectively documented the relationship between climate, conflict, and forced displacement, and created predictive models to help anticipate potential hot spots. Omdena scientists extracted data on changes to vegetation, water, and land use from satellite images, and correlated this information with UNHCR data on forced displacements and conflict. Insights from this project promote the effective allocation of resources and personnel to help the 2.6 million people currently displaced in Somalia. These findings can also inform the creation of programs and public policies aimed at reducing conflict, forced displacement, and environmental damage.
Omdena’s data scientists helped Swedish startup Spacept in their work to prevent power outages and forest fires by using satellite imagery, deep learning, and computer vision to develop a model capable of identifying the location of trees with 95% accuracy. Automating this process helps decrease the time and cost of infrastructure inspection, and reduces the risk of power outages and fires sparked by falling trees – saving lives and lowering CO2 emissions.
Omdena’s data scientists worked with the University of Bern, Switzerland to use satellite image data on Mars to classify eight types of anomalies on the Martian surface, with precision measures between 90-99%. Omdena’s scientists developed an AI tool to access, process and analyze space satellite data. Their computer vision model illustrates the state of the art for technosignature identification and helps evaluate the potential of using these approaches in future space exploration.
Omdena Satellite Imagery Case Studies 2019-2021
1. Addressing Global Poverty with Renewable Energy
|Project Partner||Renewable Africa 365 (RA365)|
|Project Goal||Address energy poverty in Nigeria by identifying those areas most suited for solar power stations.|
|Results||Built an interactive map indicating the Nigerian regions best suited for solar power installations, along with a spreadsheet ranking the opportunities.|
|Approach||Omdena’s AI community used satellite imagery, energy grid analysis, population analysis, and AI to identify where the energy poverty crisis is most dire and where solar power is most likely to be effective.|
Omdena’s team addressed challenges of incomplete and inaccurate data by correlating multiple data sources – combining information from satellite imagery and Google Earth with population data from sources such as the Demographic and Health Surveys (DHS) program, WorldPop, and GRID3. The best candidates for solar energy are mid to large-size communities that are distant from the existing power grid. Omdena’s data scientists first leveraged satellite imagery to identify those areas of the country that go completely dark at night. They then applied cluster analysis to population data to highlight the communities in those regions with mid to large populations, as well as hospitals and schools in need of stable power sources. Finally, they looked at energy grid coverage to identify the communities which are distant from the existing national electricity grid.
Expanding electricity access is an essential first step for improving education, healthcare, and local economies. The Nigerian NGO RA 365 is using the tools developed by Omdena collaborators to prioritize locations for their innovative renewable energy microgrids. This data has also been shared with Nigeria’s Renewable Energy Agency (REA), a major funding source for rural electrification projects. Insights from this project will enable data-driven investments and policies with the potential to impact the lives of millions in Nigeria. The tools and solutions are reproducible and can be leveraged in other high energy poverty regions around the world.
2. Reducing Hunger with Crop Classification
|Project Partner||UN World Food Programme (WFP)|
|Project Goal||Address food scarcity and hunger in Nepal with data driven insights on the location and growth of food staples such as rice and wheat.|
|Results||Developed a ML model to analyze publicly available satellite images, classify crops, and discern the demarcations between individual crop fields, with an accuracy reaching 99%. Insights from the process were aggregated into a guide for data scientists using satellite imagery data for agricultural applications.|
|Approach||Omdena’s data scientists enhanced low resolution open source satellite imagery, and then used neural networks on images from multiple spectral bands to identify rice and wheat crop fields with an accuracy approaching 99%.|
The project collaborators sourced publicly available images of Nepal from the European Space Agency’s Copernicus Sentinel-2 satellite, which provides 27,000 labeled and geo-referenced images, covering 13 spectral bands. The Omdena team first enhanced the Sentinel 2’s relatively low resolutions images by combining two super resolution techniques – Deep Image Prior and Decrappify. They then created an image classification model that used ResNet like deep learning neural network architecture and multispectral images for classification results with an overall accuracy of 98.69%. Additional training and data augmentation using mixup increased results to 99% accuracy.
This data provides a direct window into an essential part of the economic lives of many of the world’s rural poor. Accurate information on the location and size of the multiple small rice and wheat fields across Nepal can help the WFP improve resource allocation, accelerate the growth of staple foods, and reduce hunger in Nepal. In addition, Omdena’s guide to using satellite imagery in agricultural applications is useful for data scientists worldwide.
3. Improving Disaster Response
|Project Partner||ImpactHub Istanbul|
|Project Goal||Identify solutions to help mitigate disaster in earthquake prone Istanbul.|
|Results||Omdena team members zeroed in on the challenge of reuniting families in the aftermath of an earthquake by helping individuals identify the safest and shortest routes through a damaged city. They developed a prototype tool to map routes between two points that optimize for safety, as well as distance. Users can identify routes on a single case basis, or search for routes between multiple addresses in bulk.|
|Approach||Correlating “safety” with areas of low building density as well as wide roads and streets, Omdena’s data scientists combined data from satellite images and street maps, using image segmentation models, raster and vector maps, and other tools to identify the shortest and “safest” path from one part of Istanbul to another.|
Team members focused on one region – Istanbul’s Fatih District – to test their approach. For the first phase, Omdena’s data scientists used satellite images of Istanbul, ML image segmentation models, and tools such as mapbox to identify street and building data, including the distance between buildings, and the width of the streets. This produced a rasterized heatmap indicating safety or risk based on building density. Next, they calculated the shortest and safest routes based on vector maps and data from Open Street Map. Finally, Omdena collaborators combined the two outputs, encoding the risk value from the heatmap into the street graph. The final result allows a user to find the shortest and safest path from one place to another – one that favors open areas and wide streets and roads.
This prototype provides ImpactHub Istanbul with a useful tool to help ensure the safety of Istanbul’s residents in case of an earthquake. The methods are reproducible for other earthquake prone urban areas.
See also AI for Disaster Response: Predicting Relief During Cyclones here
4. Humanitarian Aid – Insights on Climate Change, Conflict and Forced Displacement
|Project Goal||Map the relationship between forced displacement, violent conflicts, and climate change in Somalia. Develop findings and solutions that can inform UNHCR decision making and streamline the delivery of support to people and communities in need.|
|Results||Correlating changes to vegetation, water and land use with data on internal displacement caused by conflict, Omdena’s data scientists produced tools and visualizations that provide the UNHCR with quantified, actionable insights, These findingings illustrate how climate changes such as drought or floods can lead to conflict, violence and forced displacement. They also demonstrate how the process is cyclical, where forced displacement and migrations can lead to resource strain and environmental damage that in turn fosters conflict and damages communities.|
|Approach||Omdena data scientists leveraged a range of AI approaches, including machine learning, image classifiers, neural networks, and data visualization to analyze environmental data from weather satellite images in order to document changes to land use and the environment over time. This information was then correlated with UNHCR displacement statistics from 2016-2019. Results document how climate change can influence conflict and forced displacement – as well as how conflict and forced displacement impact the environment.|
Omdena’s data scientists focused on the area around the capital city, Mogadishu, in the Banadir region of Somalia, which was highly impacted in 2016-2017. They used image classification models to extract and classify data from satellite images of vegetation, water, and buildings by district. To identify changes in land use, such as increased building density, the team leveraged data from Landsat 8 satellite images made freely available by the USGS. Spectral signatures and a random forest ML algorithm were used to classify images and register changes between “land” and “buildings” caused by increased building density. Satellite images also provided data on environmental changes, including sources such as the Vegetation Health Index(VHI), the Normalized Difference Vegetation Index (NDVI) and the Normalized Difference Water Index (NVWI), which provided information on the level of vegetation or water in a given satellite image, as well as change over time.
Results were validated by comparison with “ground truth” data on Mogadishu, extracted from an online digital Open Street Map. Collaborators additionally found that k-means clustering produces results close to those from Open Street Map, which is useful for the many rural areas of Somalia that have not been digitally mapped. Finally, Omdena’s data scientists used neural networks to correlate these land and environmental changes with UNHCR data on forced displacements and conflict.
Insights from this project promote the efficient and effective administration of UNHCR resources and personnel to help the 2.6 million people currently displaced in Somalia. Omdena’s predictive models can also help the UNHCR anticipate future hot spots, and plan accordingly. These findings can also inform the creation of programs and public policies aimed at reducing conflict, forced displacement, and environmental damage.
Read more about the article “Using Neural Networks to Predict Droughts, Floods and Conflict Displacements in Somalia” here
5. Preventing Wildfires with Tree Identification
|Project Goal||To build an AI solution to identify trees that are too close to power stations, and thus help reduce the risk of power outages and forest fires.|
|Results||Used satellite imagery, deep learning, and computer vision to build a model to identify the location of trees with 95% accuracy. This model is able to successfully identify trees in new images, even distinguishing between forest shadows and trees, exceeding the accuracy of manual labeling.|
|Approach||Challenge participants tested a variety of ML approaches for training AI algorithms to distinguish between trees and other landscape features. The best approach, a deep U-Net model, detected trees with 95% accuracy.|
The training dataset contained around 200 diverse satellite images from Australia, featuring a range of different landscapes – arid, forests, farms and urban. Omdena’s data scientists used Labelbox annotation tools to manually label the images for model training. Next, they experimented with multiple neural network models for computer vision, including Mask R-CNN and a range of U-Net variations, in order to identify and label individual trees in the images. The most successful approach – a basic Deep U-Net CNN solution – reached 94% accuracy, demonstrating the model’s ability to identify trees in new images, and even distinguish between forest shadows and trees. Additional data augmentation enabled the model to reach 95% accuracy, performing better than manual labeling.
These solutions are currently being implemented in Spacept’s product. Automating the identification of trees enables Spacept to efficiently correlate this data with power station location data to identify potential danger zones. These insights help reduce the time and cost of infrastructure inspection, and will reduce the risk of power outages and fires sparked by falling trees – saving lives and reducing CO2 emissions.
Learn more about Detecting Wildfires Using CNN Model with 95% Accuracy here
6. Advancing Space Exploration with Anomaly Detection
|Project Partner||University of Bern, Switzerland|
|Project Goal||To design a model that can detect and classify anomalies on the Martian surface, with the aim of identifying terrestrial (or extra-terrestrial) “technosignatures” – measurable properties that provide scientific evidence of past or present technology.|
|Results||Omdena collaborators created an AI tool capable of analyzing satellite images of Mars and identifying and labeling anomalies, distinguishing between 7 types of “natural anomalies” and a “terrestrial technosignature” with precision scores of 90-99%.|
|Approach||Team members created a custom Python package to efficiently process large satellite image datasets from the Mars Orbital Data Explorer. They trained and tested a variety of different ML models for identifying and classifying anomalies according to eight different classes – 7 natural and one “terrestrial” or caused by humans. A U-Net model yielded the best scores for classifying anomalies, with precision measures between 90-99%.|
Figure 6: Initial satellite image and labeled result
Omdena collaborators first created a Python package – mars-ode-data-access – to process large image datasets from the Mars Orbital Data Explorer efficiently. They then identified 8 classes of anomalies – 7 natural classes (craters, dark dunes, slope streaks, bright dunes, impact ejecta, spiders, and swiss cheese) based on a Mars orbital image dataset available on Zenodo, and one class for terrestrial technosignatures, which are the detritus or result of different Mars lander missions. Team members classified and labeled approximately 300-400 images for training. Finally, they experimented with four different ML models – 3 supervised models (SSD, Mask R-CNN, and U-Net) and 1 unsupervised model (Ano-GAN). The U-Net model yielded the best scores for classifying anomalies, with precision measures above 90 percent.
The pipeline developed by Omdena helps scientists access and analyze space satellite data. The final computer vision model illustrates the state of the art for technosignature identification and helps evaluate the potential of using these approaches in future space exploration.
Learn more about Anomaly Detection on Mars Using Deep Learning here
About the author
Dr. Karina Alexanyan is a social scientist and communications professional with 15+ years of experience managing international programs and research initiatives exploring our changing relationship to information technology.
Dr. Alexanyan’s diverse career is united by a focus on global citizenship, the impact of technological change on society, and international education. Her passion lies in understanding and leveraging emerging technologies for social benefit.