DAS predicts winter crop type within 1.1 million paddocks across mainland Australia within the growing season each year, with data available from 2019 to present.
The ten most common winter crop types are mapped and outputs from paddock to farm to regional or State scales are delivered to a range of agribusiness customers.
The accuracy of our crop classification models is obsessed over by our data science team. Because, obviously, it goes to the heart of the value and utility of our Crop ID data product.
Crop ID accuracy is quantified by using a test set of 25% of the total ground truth data, with ~15,000 ground truth points collected Nationally each year. The test set is selected randomly, but stratified by State and crop type. The test set is withheld from the ML model training process to be a truly independent sample. Overall accuracy is calculated as the number of correct predictions as a proportion of total predictions.
Overall accuracy, assessed Nationally across all crop types, of the 2021 DAS Crop ID dataset was 93.4%.
But accuracy is not uniform.
Accuracy varies spatially
State level accuracy ranged from 89-95% in 2021.
Differences in accuracy occur between regions within States
Accuracy varies between crops. 2021 per crop National accuracy was
Wheat - 95%
Barley - 89%
Canola - 97%
Oats - 89%
Chickpeas - 93%
Faba Beans - 95%
Lentils - 95%
Lupins -94%
Field Peas - 91%
Vetch - 89%
Fallow - 89%
Pasture - 94%
Accuracy varies temporally
Within the season accuracy increases as crops progress through phenological stages.
Accuracy is better some years compared to others
DAS is constantly implementing the latest machine learning methods. Each time we do, we update previous years’ Crop ID.
When you view DAS Crop predictions for any given paddock, you should expect that between 7 and 10 out of 10 paddocks are correctly identified. What you won’t see are crops other than what is listed above (e.g. triticale or rye - likely labelled as another cereal; or safflower, linseed, etc), and fodder crops may be confused with pasture or the crop type itself (e.g. fodder canola or fodder barley).
Accuracy is also constrained by the fact we only classify one crop type per paddock, and the paddock boundary data we use (“ePaddocks” produced by the CSIRO) occasionally combines multiple real paddocks into one shape. And sometimes a grower splits paddocks into two crop types. Currently, we have no mechanism to address this, but we are working on it.
If your application of DAS Crop ID data requires a more nuanced understanding of accuracy please reach out to DAS for more information. We examine accuracy every day and will openly share where we are good and where we are seeking to improve.