Image top dependent operations

[bibinwils]

I have some image datas, which are dependent on top axis of image. For example, as shown in the image, that image is dependent on the topmost line.

So I need to develop a network with always having that top line in consideration. So normal 3x3 convolutions with stride won't work in this case since by striding, convolution will come out of those top areas too. I'm trying to implement a segmentation network. Is it possible to implement any operation to always consider the topmost part?
I'm considering the topmost always, since depending on the top part of image, the entire image down rows changes. so data coming in bottom rows are an average from top to the point of consideration of bottom row.