Government & Public Sector
The determination and monitoring of agricultural areas are of immense importance for the control of subsidies and cultivation methods, e.g. fertilizers used, as well as for predictive analytics in food production. On-site inspections and manual documentation are only suitable to a limited extent due to the large agricultural areas and the time required. By automatically analyzing remote sensing data, agricultural areas can be classified at an early stage and their development over time can be monitored.
Motivation
Subsidies for agriculture are a key aspect of agricultural policy. In the EU's Common Agricultural Policy (CAP), for example, the majority of the total budget is invested in subsidies. In order to check whether the conditions for subsidies are met, data must be collected and documented on the crops grown at a parcel level. In addition, automated crop classification and subsidy control can make the documentation requirement less bureaucratic for farmers.
Furthermore, the automated analysis of agricultural land using remote sensing data can be used to monitor compliance with guidelines and the sustainability of cultivation methods, as well as to predict crop yields for the population's food supply. This enables early action to be taken and guidelines to be enforced.
Challenges
Automating the classification of crops and localization of agricultural land requires the use of earth observation and remote sensing methods and data. ML algorithms are now used as standard to analyze these large amounts of data and multispectral information. Several datasets of open source satellite data, e.g. from ESA's Sentinel satellites, can be used for this purpose. Normally, several sources of satellite data and geoinformation must be combined and standardized into one data set in order to create a suitable training data set. This must also include labeled and annotated crop types and growth phases as well as other vegetation types.
Solution approaches
Open-source datasets, such as the data from ESA's Sentinel-1 and Sentinel-2 programs, provide optical and radar-based satellite data and thus various spectral information about the Earth's surface. ML algorithms are able to simultaneously process this different information and features to identify crops.
Image segmentation algorithms, which are usually Convolutional Neural Networks (CNN), are implemented to classify plant species. They can identify differences in color, growth height and moisture of crops and other vegetation via spectral information and classify crop species and the boundaries between different fields and forests on this basis.
In order to extract additional information from different growth phases and observe the temporal development of the areas, time series of satellite images are used and the time-dependent relations are analyzed using transformer models.
This makes it possible, for example, to determine the type of crop grown on a field early in the year and to check whether the conditions for the approval of subsidies have been met or whether the information provided by the farmer is correct.
