CropID’s machine learning approach integrates multi-spectral satellite imagery (Deimos 1 and UK DMC 2 22m resolution) soil properties and physical field characteristics. Image processing algorithms are used to segment crops based on spectral reflectance and texture, based on a small ground truth survey. Satellite images acquired through the year allow the system to build knowledge about the crops in individual fields. Soil properties are used to focus the analysis on areas suitable for specific crops.
The benefits of using the remote sensing approach of CropID to support the production of agriculture land use statistics are:
- Speed – delivering a crop map within the current growing season
- Low cost – classification is achieved using limited ground data and using more automation in the processing chain
- Accuracy – results are more reliable than current sampling and statistical analysis techniques
- Accessibility – crop data is easy to visualise and explore
- Perspective – provides new insights for farmers, growers, researchers and policy makers
The classification of horticultural crops is complex, especially taking into account the huge number of different crops that exist and the fact that some types of crop might have a very similar “signature”. The seasonal growth cycles of vegetation (phenology) can be used to help distinguish vegetation types, potentially individual crop types.
Determining the timing of key growth stages, i.e. emergence, canopy development, maturation, senescence, can be useful to aid classification as different species emerge and mature at varying rates. However, a key requirement for a satellite platform to differentiate crops based on their phenological characteristics is a high temporal resolution (revisit time).
