Making digital maps of Fields, usually made in expert mode, using space images or aerial photography data, including unmanned aerial vehicles, or by detour path with specialised geodetic equipment. It is quite labour intensive and expensive process. Besides, making maps for huge area, for example, in regional scale need a lot of time.

We introduce product, based on modern technologies of machine learning and computer vision, using process of space images at different times and programs Sentinels-2. It made substantially possible reduce the spending of time on creating field maps, consequently to decrease cost of the product.

The methodology, which is used by us is based on the object-oriented approach, widely used in automated decoding of space images, consisting of two main stages - segmentation and classification.

Segmentation - combining the original image pixels into adjacent areas, the so-called super-pixels. There are several generally accepted approaches for obtaining super-pixels, in our case, this is the WaterShed algorithm, while the maximum values of the NDVI vegetation index gradient are used as the boundary condition (the maximum of the function), and the values of the NDVI index are used as the marker (minimum of the function) corresponding to the open soil (less than 0.2).

To ensure the uniformity of segmentation (the maximum possible coverage of the processed region with segments), this process is performed iteratively over a series of satellite images at different times, made for two agricultural seasons.

Classification - assignment getting from segments to the target class (arable lands). Super-pixels obtained at the stage of segmentation can be formed both by arable lands and by other natural or man-made objects, for example, water surface, buildings and roads, quarries, as well as by objects located under the shadows of clouds. A classifier is used to distinguish among all the obtained segments, those that relate to arable lands. In our case, this is a fully connected feedforward neural network (MLP) trained on a large (more than 10 thousand) set of hand-labeled examples. The following are used as classification signs: the dynamics of changes in the NDVI and NDWI index, the value of the gradient along the border of the segment, the spread of the values of the indices and the gradient, characteristics of the shape of the segment.

As a result, each obtained segment is assigned class 1 or 0 (arable land or other), and the value of the probability of assigning this segment to the class. Then, vectorization is carried out - the translation of segments to a vector form of representation, for the possibility of analysis and display in GIS packages and web services.

The attribute information of a vector layer contains information about the probability of assigning each contour to arable land, which allows data analysis at various levels of significance.

Below you can see examples of processing results for the Saratov region located in Russian Federation