Context and objectives
Crop and yield forecasting is key for agricultural cooperatives. Knowing harvested quantities as early as possible, at a fine geographical level, lets them optimise their supply chain, especially at harvest time, when tonnes of crops must be processed, transported and stored in silos.
Outcome
The harvest prediction models are used to initialise and anticipate the harvest plan, which now optimises the supply chain, with at least a 10% reduction in transport costs and CO2 emissions.
Our approach
Step 1: reconstruct
- Reconstruction, with agricultural cooperatives, of a unique, large database of fine-grained historical crop types and yields
Step 2: retrieve
- Retrieval and processing of satellite images to build vegetation indices (NDVI). Combined with strong forecasting expertise, these fine-grained spatial and temporal indices make it possible to link them to crop types and yields.
Step 3: build
- Machine learning models that estimate the crop type and final yield several months in advance, with an error below 17% for more than 86% of agricultural parcels