4m read

Predicting crop yields with satellite imagery

4m read

Inference is a visual analytics platform that provides a real-time view of global crops

Succeeding in global commodity markets, requires robust data that gives you an informational edge. But all too often, the available data is patchy and unreliable.

Parallel partnered with HSAT – a leading machine learning company – to design and build a visual analytics platform that aims to increase coverage and reduce uncertainty, using automated analysis of satellite imagery and ground level photography.

The challenge

In order to provide their clients with meaningful insights, HSAT are using powerful machine learning algorithms on vast amounts of raw data sourced from satellites such as the European Space Agency's Sentinel-2.

In addition to the raw data being hard to work with, the standard applications typically used to analyse this type of data are complicated tools that require academic-level expertise. Given that HSAT's clients operate in a fast paced environment and are not domain experts in geospatial analysis, it was vital that the Inference platform delivered insights that were focused, timely and trusted – with the ability to validate findings using additional sources of data.

Our approach

Sweat the important stuff
Our primary design principle in response to this challenge was to focus on the key metrics and functionality that commodity traders need - and to strip out everything else to reduce the time-to-insight ratio and provide bespoke and highly-processed data points.

The first crop HSAT focused on was sugarcane grown in Thailand. The country is the third largest exporter of sugar in the world, but the reliability of existing data about both the quantity and quality of the crop is often sub-optimal. This is due to the highly fragmented agricultural structure, where farms are often family run businesses that tend to rotate crops far more regularly than in other big producing countries such as Brazil or India.

For someone looking for commodities intelligence in this sector, the most important metrics are the area of land given over to sugar cultivation, levels of precipitation, yield and price over time.

The application was design to provide two main views on this data. The first showing historical data on sugarcane volumes (hectares) and precipitation (highly correlated to yield).

Balancing macro and micro
For the Thai sugarcane use case, traders and analysts typically want to understand patterns at a macro level - the expected volumes and yields for entire regions and provinces. The second view of the application provides this through heat-map visualisations that instantly direct the viewer to the provinces with the highest amount of sugar crops as well as the biggest changes in sugar across a number of years.

Users are also able to view heat-maps for total cropland and yields. Both metrics serve as indicators for future growth and sugarcane quality.

Alongside the aggregated heat-maps, Inference offers much higher resolution imagery as you zoom into the maps. A key visualisation here were the lidar-like 'masks' that show individual fields in which the computer vision system has detected sugarcane is being grown.

Reducing uncertainty
It was also important that users were able to trust the data. As with most new technologies, it is prudent to maintain a level of oversight and understanding about what machine learning algorithms are seeing and interpreting. HSAT's approach to this is to periodically enhance their satellite imagery datasets with the ground level imagery. This is done by sending out a mix of scouts and drones that will probe the crop lands with both video footage and photographic images.

This approach has two benefits: not only does it help human analysts to validate the data that they are seeing on the platform - via editorial style reports - but it also serves as a feedback loop for the detection algorithms, which improves their accuracy over time.

Scaling up - Platform strategy

HSAT are already providing intelligence and functionality for a variety of crops from different regions around the world. To enable the platform to scale to meet the various needs of their clients, we adopted a matrix approach to their offering. Clients can configure their subscriptions to include multiple functional components and specific locations around the world.

Additional components include Field Viewer to help users assess the likelihood of crops suffering from viruses and Contract Manager to provide detailed information of the crops being grown and expected yields for specific farms within a portfolio.

Insight from above

From crop intelligence to monitoring emissions and battling deforestation, the emerging digital twin of the earth, enabled by aerial and orbital imaging will provide us with entirely new ways to observe, analyse and predict our world. But to make a difference, the information hidden within these images needs to be seen - and understood - by the right people. The domains they operate in will be very specific and the data they need highly-dimensional.

Making the most of today's advanced analytics techniques requires intuitive, domain specific applications that can integrate multiple data sources and playback meaningful insights. If you have an advanced analytics application you'd like to discuss, please reach out to hem@parallel.systems