For insurance purposes, estimating the value and the risk of a building is not an easy task. An expert has to go to the location and assess the house in person… I wouldn’t trust photos taken by the owner! This not only implies time and travel costs, and it could be even impossible during a pandemic, with lockdowns and limitation to the movement of people.
There is where Remote Sensing and Artificial Intelligence come into play. Images could be taken by satellites, airplanes or drones and then processed and analyzed by Neural Networks.
Data scientists at Balzano worked on a project whose goal was to estimate the roof area of houses, classify the roof material (wood, metal, tiles, thatch or concrete), detect the presence of trees nearby, which increases the risk (of fire, for instance) and determine if the property had a swimming pool, which obviously increases the value.
Images came from satellites: high resolution gray scale images, low resolution colored images. We combined them to create high resolution colored images:
A tool for cropping the huge satellite images into smaller pieces and for segmenting the roof area from the background had been internally developed:
A Neural Network, whose inputs were Images and Masks (manually drawn), was trained to identify the roof and compute its area:
The network architecture was based on U-net. The network consists of a contracting path (the encoder) and an expansive path (the decoder), which gives it the u-shaped architecture. The contracting path is a typical convolutional network that consists of repeated application of convolutions, each followed by a rectified linear unit (ReLU) and a max pooling operation. During the contraction, the spatial information is reduced while feature information is increased. The expansive pathway combines the feature and spatial information through a sequence of up-convolutions and concatenations with high-resolution features from the contracting path, thanks to “skip connections” from encoder layers to decoder layers that are on the same “level”.
Here are some results: even in complex cases, where the shadow of the houses could have fooled the algorithm, the output was quite good.
Here is an outline of our solution. We integrated Deep Learning with traditional Computer Vision algorithms for improving the results where the neural network was not accurate enough, during pre- and post-processing.
Artificial Intelligence can help analyze images taken by satellites, or even drones, in many different ways. Not only can it be used to evaluate the value and risk of buildings, like in this project, but it can also help to evaluate damage after flooding or fire, aid in Catastrophe Management, estimate the status of roads and bridges or, thinking more positively, enable “precision agriculture”: https://en.wikipedia.org/wiki/Precision_agriculture.