Semi-automatic mapping of charcoal kilns from airborne laser scanning data using deep learning

This paper proposes the use of deep learning for semi-automatic mapping of charcoal kilns from airborne laser scanning data.
A deep convolutional neural network (CNN) was first pre-trained on 1.2 million photographs in order for the network to learn
general high-level image features. The second to last CNN layer was input to a linear support vector machine, which was trained
from CNN features obtained from 375 charcoal kiln locations and 10,000 other locations.
In a 3 km × 3 km test area, the automatic method identified 363 of 419 verified charcoal kilns, while 56 were missed. Nine
previously overlooked, possible charcoal kilns were also found. The number of false positives was 220.
The proposed method, based on deep learning, is better than our previous attempts at semi-automatic charcoal kiln detection
based on traditional pattern recognition methods. The new method detects more true charcoal kilns and has a manageable
number of false positives.