Using EO to support sustainable agriculture practices in Brazil
Norman Kiesslich, GeoVille GmbH, Austria
The objective of this Crop Explorer activity with the ODA organization GIZ and the Central Bank of Brazil (BCB) was to assess compliance with sustainable agricultural loan requirements by detecting crop type, sowing and harvest dates and comparing that to information provided by the loan applicants. In total 70.000 land plots in Brazil that were checked for cultivation of the three main crop types wheat, corn, and soy. Furthermore, the percentage of the area of the field that is under agricultural practise was assessed and reported for each parcel, since the claimed field outlines do not necessarily contain only crops but also other land cover like forest, water bodies or roads. Confidence values for the detected crop type are provided for each plot. The provided results put the bank in a position to specifically address clear violations in terms of false crop types cultivated or very significant deviations from the reported sowing and/or harvest dates.
For this activity, BCB provided several thousand polygons of various known crops that were serving as sample data for the training of the crop detection algorithm. The machine learning algorithm was fed with full spectral Sentinel-2 timelines for each land parcel spanning from 30 days before the indicated sowing and 250 days after declared harvest. Within this timespan, the NDVI timeseries and other parameters are analysed and trained for the application of the dataset that needs to be checked. The dataset shows a very high accuracy of detecting the correct crop type, which lay between 96% and 99%, depending on the target crop.
The algorithm calculates the percentage describing the probability of every crop type on a specific field. The image below shows example results, as displayed in the CropExplorer, for the three cases of detection classification. If the algorithm is detecting one crop type with more than 80% certainty, it is declared as the detected crop type and classified as correct if it matches with the declared crop type (see Soja) or as wrong if it does not match (see Milho). If no crop type can be detected with more than 80% certainty, it is classified as suspicious (see Trigo) and needs further investigation.
Benefits & next steps
The ability to detect individual violations from tens of thousands of applications without the need for on-site inspections is a tremendous advantage over the current status quo and a clear benefit for the stakeholders. The BCB thus provided an overly positive feedback regarding the preliminary results of the current activity. It is noted that the scale of the challenge for a region as large as Brazil is significant and talks are ongoing to expand this activity to other regions of Brazil as well as additional crops. We will update on progress within this project in due course.
If you would like to know more please contact us via the email@example.com