When a convolutional neural network downsamples the image representation as it passes through its layers, the value of the feature maps at a given location p in a given layer L is the result of operations on pixels in a larger region of the previous layer, L-1, and an even larger region of the layer previous to that, L-2. Proceeding as such all the way to layer 1 (the input image) we specify a certain region on the first layer which influences the value at p. That is the receptive field of the location p in the layer L.