Proceedings Article | 6 July 2022
KEYWORDS: Satellite imaging, Neural networks, Earth observing sensors, Unmanned aerial vehicles, Remote sensing, Convolutional neural networks, Soil science, Raster graphics, Agriculture, Satellites, Satellite imaging, Neural networks, Earth observing sensors, Unmanned aerial vehicles, Remote sensing, Convolutional neural networks, Soil science, Raster graphics, Agriculture
Soil erosion all over the world is an intensive, poorly controlled process. In many ways, this is a consequence of the lack of up-to-date high-resolution erosion maps. All over the world, the problem of information insufficiency is solved in different ways, mainly point-by-point, in local territories. Extrapolation of locally obtained results to a larger area inevitably leads to uncertainties and errors. For the anthropogenically developed part of Russia, this problem is particularly relevant, because the assessment of the intensity of erosion processes, even with the use of erosion models, does not allow to achieve the required scale due to the lack of all the necessary global large-scale remote sensing data of the Earth and the complexity of considering regional features of erosion over such large areas. The paper proposes a new technique for automated large-scale mapping of erosion processes, namely rill erosion, according to Sentinel-2. Deep learning neural networks are used to recognize washouts. To recognize rills, a transfer learning approach was used, namely, a combination of the LinkNet architecture with the EfficientNetB3 encoder. The accuracy of automated recognition of rill erosion in the study area of 3,200 square kilometers is 80% compared to the results of manual recognition. The average density of the rill erosion was 0.3 km/sq.km, the maximum – 0.66 km/sq.km. The greatest density of washouts corresponds to the plowed deforested territories actively used in agriculture, the minimum – is on cultivated lands with contour farming.
