Changes

So basically we need an Image to Image neural network that recognizes the parts of the agricultural fields black and the rest white (or the opposite).

To simplify the training of the network it could be a good idea to develop an iterative image recognition algorithm, based on color, texture, ... first and use it to provide learning data with google maps. The automatically elaborated pictures can then just be corrected with human interaction (smooth workflow needed!) to provide the needed gain in quality.