What are Siamese Networks, How Do They Work and how to apply them to mineral exploration?
Siamese Networks (SN) are an AI solution within neural network architecture, designed for tasks that involve comparing two data inputs to determine their similarity or difference. The paired structure allows the network to learn a “similarity metric,” which is especially useful in scenarios where subtle differences between input data are essential. In image processing, for instance, Siamese Networks are applied in facial recognition and signature verification tasks, where distinguishing fine-grained details is crucial. In the context of mineral detection, Siamese Networks have shown remarkable potential. Their architecture makes them adept at recognizing complex patterns in hyperspectral images, which contain plenty of spectral data across a specific for the product range of wavelengths. By analyzing spectral signatures—unique patterns of light absorption and reflection distinct for each mineral—Siamese Networks can learn and detect specific minerals with greater precision than traditional methods, including minerals with closely related spectral profiles. Unlike other methods utilized in TerraEye, for example the Spectral Angle Mapper (SAM), which relies on angle calculations between database spectra and spectra derived from the satellite image, Siamese Networks use deep learning to identify even subtle spectral variations, allowing them to excel in identifying minerals within complex mixtures or, in other words, densely mineralized environments.
From a geological perspective, the output of a Siamese Network offers valuable insights for mineral exploration by generating similarity scores that can be mapped to show mineral distribution. The network calculates a “distance” metric between spectral data collected for a specific area of interest and the datasets from spectral libraries (for example the USGS spectral library); the smaller the distance, the higher the similarity between those two inputs. For example, SN detects a field spectrum which has short metric “distance” to a library spectrum of alunite and thus will classify the field spectrum as alunite. This information helps geologists and exploration teams understand the mineralogical composition of target areas and assess the presence and concentration of valuable resources.
Implementing
Siamese Networks in TerraEye's Workflow
In TerraEye’s use
cases, Siamese networks are used as a tool to advance mineral detection
capabilities provided by methods which are already implemented in the TerraEye
app. SN provided exceptional results especially within hyperspectral imaging
for geological exploration, as hyperspectral imagery provides the highest level
of spectral detail.
By training the network
on a dataset of known mineral spectra, TerraEye develops a model that detects
and maps minerals accurately, even when multiple minerals (mineral mixtures)
are present in the same pixel. This capacity to continuously improve by learning
from various product datasets makes Siamese networks a valuable asset for
TerraEye, allowing us to deliver highly accurate mineral maps for mineral
prospecting and exploration.
app.terraeye.co
www.terraeye.co
