Projects for Autumn Quarter 2018-2019
1) Africa crop yield forecasting
To predict subnational crop production based on weather + satellite data, following approach used in last year's project in South America. Useful for a range of applications including early warning of areas with shortfall in production.
2) Humanitarian tent detection
To establish methods with high recall and precision in determining number of tents in refugee camps, based on historical imagery with manual labels. Useful for agencies that need to estimate resource needs for aid on a weekly basis in areas where camp sizes often grow and shrink rapidly.
3) Crop type mapping
To combine high temporal frequency (weekly) radar and optical data to improve ability to assess which crops are growing across landscapes in Africa. Useful for World Food Program in assessing capacity for local production in conflict-prone areas like South Sudan. Also useful for other applications around understanding crop productivity or land use decisions by farmers.
4) Famine intensity prediction
To predict variation in the IFC classification score for famine (intensity scale from 1-5) at level 2 admin level based on weather, modis imagery, food prices, night lights. Would be useful for agencies tasked with responding to and helping to avoid famines.
5) Global poverty mapping with Wikipedia
To predict outcomes related to wealth and infrastructure based on using publicly availalable wikipedia data. The approach could include using features extracted directly from wikipedia posts, or using features or object counts from image-based models that have been trained on classification tasks for wikipedia-based labels. Would be useful for mapping baseline levels and changes in key development outcomes in remote areas.
6) Air quality
To track changes in air quality from photos taken from webcams at regular intervals, building on work from a prior group project. Would be useful for a variety of studies to understand health impacts of air quality and more cheaply monitor changes over time.
Projects for Winter Quarter 2018
1) Mapping road and building infrastructure
This project will test the ability of deep learning models to map roads in developing countries based on moderate (10-30m) resolution optical and radar imagery (Landsat, Sentinel-1 and Sentinel-2). Training data on roads and building density will come from OpenStreetMap in Africa and Asia. As an additional task, features extracted from these models will be tested for their ability to predict local road quality as derived from field surveys.
2) Forecasting food prices in India
This project will test the ability of geolocated tweets from India to track fluctuations in the prices of major food staples in India. The core datasets will be 2.5 years of tweets obtained as part of Stanford's Data Science Initiative, and local weekly price data in India from the World Food Program. Past work has suggested people discuss food more when prices are rising, such as in an Indonesia study by the UN Pulse Lab, but the concept has not been widely proven. The goal is to produce timely warnings of where prices are changing, particularly if they are moving very rapidly, as a way for governments, NGOs, and the private sector to cheaply monitor and respond to these situations.
3) Tracking displaced peoples in humanitarian crises
This project will test the ability of high-resolution (<1m) satellite images to map the density of tents in humanitarian crises. An ability to automatically track displaced people would reduce the amount of manual image labeling done in humanitarian organizations, and lead to more efficient distribution of scarce resources. Training data will be provided from partners at World Food Program and UNHCR.
4) Tracking changes in air quality around the world
Data on air quality is poor in many parts of the world, despite the fact that it is a major health hazard. This project will test the ability of photos taken from webcams at regular intervals to track changes in air quality, as measured by on-the-ground monitors in the US and Europe. Baseline models will include using existing metrics used for photo haze correction, which will then be compared to more sophisticated deep-learning models.
5) Mapping within-city variations in infrastructure
This project will test the ability of high-resolution (<1m) data to map important aspects of urban infrastructure (e.g., water access, garbage removal, quality roads) within Addis Ababa, Ethiopia. The core datasets will be multiple Skysat images taken during 2016 as well as detailed household surveys taken throughout the city. An ability to accurately map current infrastructure for large cities would help governments allocate scarce resources, and an ability to track changes over time would provide an effective way to measure progress.
6) Mapping specific types of facilities
This project will test the ability of fine and medium resolution data to track the presence of objects that match those in a training set of images. (see Geovisual search for an example). The two specific tasks for this project will be to map (i) brick kilns, which are a major source of local pollution, in South Asia and (ii) confined animal feedlot operations (CAFOs) in the United States. Better data on the location of these facilities will help to better understand effects on human health and agriculture in the vicinity. Training data will include locations collected by Stanford colleagues.
7) Mapping soil quality
This project will test the ability of multitemporal moderate (10-30m) and coarse (1km) resolution data to map specific soil properties. Training data will include thousands of soil samples recently collected by colleagues in India. An ability to accurately map soil features would aid the targeting of agricultural technologies better suited to specific soils.
8) Mapping poverty in India and Bangladesh
This project will build upon past work that mapped poverty using CNNs and high resolution imagery (Jean et al. 2016). Two new datasets that include household-level measures of assets and expenditures will allow further refinement and testing of past approaches. In addition, the team will use new sources of imagery, including Sentinel-1 radar data, that could be useful for poverty prediction. The goal is to produce reliable maps that can be updated over time, in order to track the progress of communities in building assets and wealth, and test hypotheses about which factors speed up or slow down progress.
Projects for Fall Quarter 2017
1) Mapping infrastructure in Africa
This project will test the ability of deep learning models that use a combination of high (~1-3m) and moderate (10-30m) resolution optical and radar imagery to predict measures of infrastructure in Africa. Training data on measures such as access to electricity, quality roads, and piped water will be from the recently georeferenced Afrobaromter surveys of multiple countries, as well as detailed field data from Addis Ababa. The goal is to produce reliable maps that can be updated over time to track the provision of basic public services.
2) Mapping poverty in Uganda, Bangladesh, and India
This project will build upon past work that mapped poverty using CNNs and high resolution imagery (Jean et al. 2016). Three new datasets that include household-level measures of assets and expenditures will allow further refinement and testing of past approaches. In addition, the team will use new sources of imagery, including Sentinel-1 radar data, that could be useful for poverty prediction. The goal is to produce reliable maps that can be updated over time, in order to track the progress of communities in building assets and wealth, and test hypotheses about which factors speed up or slow down progress.
3) Forecasting crop production around the world (esp. Africa, Latin America)
This project will use primarily satellite data from MODIS (both surface reflectance and temperature) with CNNs and Gaussian processes to forecast crop yields. This approach was first developed using U.S. data for soybean and maize in You et al. (2017). This project will start with that model and then extend it for application to sub-national crop datasets in Argentina and for several countries in Africa. The goal is to produce accurate estimates of final yield at various lead times, from several months before to the month of harvest.
4) Mapping land cover around the world
This project will develop methods to map the occurrence of cultivated croplands around the world at high spatial resolution. The core dataset will be ~50,000 high resolution images with crowdsourced labels of whether or not cropland is present in the image, as well as coarser 10m resolution images from Sentinel-2. The first step will be to see whether deep learning models can reproduce the human labels on the high-res imagery, and the second step to see whether the 10m data work nearly as well. The goal would be to use the 10m data, which is available for free globally, to produce a global map of where crops are. This information would be useful for a wide range of applications, including developing a mask to apply to more sophisticated analyses of crop yield (such as in project #3).