Estimating Elevation via Deep Learning

A deep learning model for estimating construction site elevations using a drone-based orthoimage.

The proposed method includes an orthoimage-based convolutional neural network (CNN) encoder, an elevation map CNN decoder, and an overlapping orthoimage disassembling and elevation map assembling algorithm.

In the convolutional encoder-decoder network model, the max pooling and up-sampling layers link the orthoimage pixel and elevation map pixel in the same coordinate.

The experiment data sets are eight orthoimage and elevation map pairs (1,536 × 1,536 pixels), which are cropped into 64,800 patch pairs (128 × 128 pixels).

Experimental results indicated that the 128 × 128-pixel patch had the best model prediction performance.

After 100 training epochs, 21.22% of the selected 2,304 points from the testing data set were exactly matched with their ground truth elevation values; and 52.43% points were accurately matched in +/-5 cm and 66.15% points in +/-10 cm, less than 10% points exceeded +/-25 cm.