What's wrong with the neural network?

Question

Neural networks

Good afternoon, I'll try to be brief.

I'm trying to analyze a satellite image for forest/field classification. I took a picture from a range where it is obvious where the field is and where the forest is (one light, the other dark and very contrasting), set 1 and 2 for classification, drive one value - color from 2000 to 5000, that is, nowhere is easier. Two fully connected layers and a 0.2 dropout. One entry, one exit. As a result, it trains for values ​​in the middle of the range, i.e. 1.5 (between 1 and 2). If you drive all 104 ranges (8 photos of 13 ranges) it learns, but does not fit. Simplified to understand why, since the classification is so simple that the eye can see. I took both a point and an area of ​​​​points 5x5, 9x9 with averaging - the same thing - 1.5 as a common value for everything and a lattice of pixels at the boundaries of this transition (I round the values ​​​​to integers after the predict).

What to do and what kind of garbage is this? Why can't it give out 1 for the range of 2000-3000 and 2 for 3000-5000 relatively speaking, if everything is visible with the eyes.