Enhancement of coarse resolution imagery using super-resolution and convolution neural networks to be used for fit-for-purpose systematic land titling and land valuation
A method of enhancing digital images of a parcel of land that includes providing a pair of coarse resolution and fine resolution orthophotos of a parcel of land to a model, training the model based on the orthophotos to obtain a trained model, and providing a coarse resolution image of another parcel of land to the trained model to obtain a fine resolution of another parcel of land.
1 . A method of enhancing digital images of a parcel of land, comprising:
a) training a super resolution conventional neural network model to improve resolution of a digital picture,
wherein the training includes:
providing a pair of orthophotos, where a first orthophoto is a coarse resolution orthophoto having a first resolution and a second orthophoto is a fine resolution orthophoto of a parcel of land having a second resolution, wherein the second resolution is higher than the first resolution, and
wherein the the first orthophoto and the second orthophoto are aerial images covering an exact same ground area, and
the second resolution is between 5 and 25 centimeters per pixel;
geometrically correcting the the first orthophoto and the second orthophoto to have a uniform scale;
training the model based on the corrected orthophotos to obtain a trained model; and
b) providing a coarse resolution image of another parcel of land to the trained model as an input to obtain a fine resolution image of the another parcel of land as an output.
2 . The method of claim 1 , wherein the first orthophoto and second orthophoto are from a same time or are from areas that have not had significant changes over time.
3 . The method of claim 1 , wherein the training step further comprises dividing each of the first orthophoto and the second orthophoto into 256 by 256 pixel images, wherein each division of the first orthophoto corresponds to a division of the second orthophoto to be used as a ground truth.
4 . The method of claim 1 , further comprising augmenting the first orthophoto and the second orthophoto prior to the providing step.
5 . The method of claim 4 , wherein the augmenting step includes flipping or randomly adding noise to the orthophotos.
6 . The method of claim 4 , wherein the augmenting step includes changing the first orthophoto and the second orthophoto by adding artifacts to the first orthophoto and the second orthophoto or obscuring parts of the first orthophoto and the second orthophoto with random text.
7 . The method of claim 1 , further comprising using learnable blur, self-attention, discriminative learning rates or progressive resizing in the training step.
8 . The method of claim 1 , further comprising using a perceptual loss function in the training step.
9 . The method of claim 1 , further comprising identifying improvements in the fine resolution image of the another parcel of land by applying semantic augmentation and image classification to the fine resolution image of the another parcel of land.