SPbGAU Student Teaches Neural Network to Determine Productivity of Farm Field Plots
The method is based on satellite imagery analysis.
Omar Tsukaev, a student at St. Petersburg State Agrarian University, is developing a new method for assessing field productivity. The approach makes it possible to analyze individual plots and build unified digital maps from that data.
A neural network is at the core of the entire development. It cross-references satellite images with imagery obtained from mapping systems on combine harvesters. To train the model, sequences of optical indices calculated for each image pixel over a given period are used. The data are normalized to a single scale from 0 to 1. The indices characterize various properties of vegetation and the field.
During training, the model receives index sequences for image pixels along with the corresponding values from the yield map. Once trained, the neural network will be able to calculate a productivity score from 0 to 1 for each pixel. The size of an elementary plot will correspond to the pixel area, that is, the resolution of the satellite image. Based on the calculated values, the field can be divided into zones with different productivity levels and displayed on a digital map.
