Skip to main content

This online seminar includes industry best practices for using UX-Analyze

to process advanced electromagnetic sensor data to classify unexploded ordnance targets. An overview of UX-Analyze, along with practical tips for experienced users to help improve their workflows.



Darren Mortimer
Product Owner – Seequent


58 min

See more on demand videos


Find out more about Oasis montaj

Learn more

Video Transcript

<v Darren>Hi everyone.</v>

My name is Darren Mortimer

and I’m a product owner here at Seequent.

On behalf of Seequent

I’d like to welcome you to today’s webinar

on industry Best Practice’s

for Advanced Geophysical Classification of UXO Survey Data.

So here’s what we’re going to cover.

What is classification

and why consider using it for your UXO projects?

And an introduction to Advanced Geophysical Classification,

along with working with dynamic and static survey data.

I also have some tips

of features that you may not be aware of.

Time savers to make your project workflows more efficient.

So whether you’re a new user to Oasis montaj

and UX-Analyze, or a seasoned pro,

I have something for everyone.

So, let’s get started.

So what is classification?

It’s the action or process of classifying something

according to shared qualities or characteristics.

Anyone can do classification.

In fact, we learned to do this a quite at an early age.

I went and found some experts

and see how well they would do.

You can see they did a pretty good job

of being able to classify or group the items

based on their property.

Things like size, shape and color.

However our classification problems aren’t quite so easy.

We need to find things like UXOs, unexploded ordinance

or ERW, explosive remnants of war.

And we must look in places like fields and forests,

where they’re not easily visible.

Now why would we want to do classification?

Several years ago, the defense science board did a study

on the typical pro of cost breakdowns

for munitions projects.

And the typical munitions clean up,

an overwhelming fraction of the money

is spent removing non-hazardous items.

So if we can save money,

if we can identify these items beforehand

and either remove them with fewer safety precautions

or simply leave them in the ground.

Another way to think about this

is if we can reduce the digging scrap or clutter by 90%,

we can see a reduction in project costs.

I should note there are sites

where classification isn’t recommended.

If you’re working on heavily impacted areas

and looking for small items,

or when you know you’re going to need to dig everything up

because of the nature of the final land use of the site.

So what is Advanced Geophysical Classification?

It’s using a principled physics-based approach

to reliably characterize the source

of a geophysical anomaly as either a target of interest,

a UXO or as a non target of interest,

clutter, debris or scrap.

And you must recognize that even the current

survey or field methods

already involve some kind of implicit discrimination.

Mag and flag, how sensitive is the instrument that’s using

and how attentive is that human that’s

working and listening to the tones

and reading the dial as they go along.

Or in digital geophysics when we set our target thresholds.

Above this, we will pick it and call it an anomaly,

below that, we don’t.

Those themselves are some levels of classification.

We found that electromagnetic geophysical methods

are the most useful.

Compared to magnetic methods,

EM is minimally affected by magnetic soils

and can detect both ferrous and non-ferrous items,

and also provides more information

or properties about the source.

Things like distance, orientation, it’s size and shape,

material type and thickness.

Some of these can be called extrinsic properties.

They’re external to the item, other one they’re intrinsic.

These are the properties that are the most important ones,

because then we can look at these

and use those for classification.

The EM response can be decomposed into components

along three orthogonal principal directions.

These magnetic polarizabilities are specific responses

to the EM excitation or the electromagnetic excitation

along the target’s or source’s principal axes.

Basically these things called polarizabilities,

completely describe the EM response of the target

and the are intrinsic to the target.

And we’ll see a little bit more about that coming up.

So thinking about some of the conventional sensors

which you may be familiar with.

These are really good for the detection,

but the not usually good for classification.

They have a limited number of measurements.

Often only a couple of time gates or maybe even one.

And the generally usually a single monostatic transmitter

and receiver.

That means the transmitter and receiver

are pointing in the same direction

and they’re at the same location.

To be able to get a full look of the target,

we need to move the sensor around

and even small errors and locating the sensor creates noise.

The end result of all of this,

that the sensors aren’t good for classification because they

don’t allow us to generate good, accurate,

reliable polarizabilities.

So along comes the advanced electromagnetic sensors.

These guys are designed for classification.

They observe the response

and allow us to calculate reliable polarizabilities.

And there kind of is two types of flavors.

There’s either a single axis planar array

where we just have a array of coils,

very similar to what you’re working with already.

Or we can mount these in

the transmit and receiver coils

in many different orientations and directions.

So it’s both allowing us to fully illuminate

or excite the target and measure it from several directions.

Here are some examples of current system sensors

that are available and are in use today.

Things like the TEM two by two

and the MetalMapper two by two,

they have four transmitter coils and a planer array.

And in the center of each of those coils is a receiver

which is a multi-axis receiver.

It is orientated in both X, Y and Z.

On the other hand, you’ve got things like the MetalMapper

and the Man-Portable-Vector or NPV.

These guys have multiple access transmitters,

and you can see them there sticking up above

looking kind of like an egg beater along with

in the case of the MetalMapper seven multi-axis receivers

and the case of the MPV, five multi-axis receivers

on that sort of a round circular head.

We also record over a much larger window.

Typically in a static survey mode,

we record 122 gates over 25 milliseconds,

collecting much more data.

And we can use this data to help us to determine

or develop these intrinsic properties of the source.

We can take our response data here which has shown

all the responses from a two by two sensor.

The plots are shown with a log time and along the x-axis,

and it’s the log voltage along the y-axis.

We can take all of this response data and invert it

to give us reliable polarizabilities.

And I have a little example here for you.

Here we have a gain that a two by two type system.

Each of the large squares represents the transmitter

in a planar array.

In the center of that there is a receiver

that has got the three receiver coils on it

in each of the three orthogonal directions.

We have the response data

and then there’s the polarizabilities.

And if I take this source

and we’re going to move it around here,

you’ll be able to see how changing the source locations

changes the response,

but the polarizabilities essentially remain the same.

So we can move it down there to the bottom

and then move it over to the top.

I’ll just go back and forth there where you can see

how the response keeps changing,

but the polarizabilities essentially don’t.

So we can use these polarizabilities

since they completely describe the EM response of the source

and they’re intrinsic to the source

and they really don’t change due to the depth

that we will bury the source or its orientation.

We can also extract from them

a number of properties which are directly related

to the physical properties of the source.

We can look at the decay rate

which will give us the wall thickness.

We can look at the relative magnitude

of the various polarizability that gives us

an indication of the shape of the item.

And we can also look at the total magnitude

of the polarizability and that will give us an indication

of the overall volume or size of the object or source.

These features or properties can be easily shown

in a feature space plot.

For example here’s the size and decay.

Remember size is kind of the overall volume of the object

and decay is that notion of the wall thickness.

And when we can use that to classify items.

Well, we can see here that we’ve got a grouping of

targets or sources there related to 75 millimeters

and other ones related to a 37-millimeter,

but the 57 millimeters, they’re a little spread out.

It’s not quite as helpful.

These feature plots or the features alone

have a limited classification power

compared to the overall curve.

These are really what the source looks like in the EM sense.

So we could compare the polarizabilities,

the whole entire curve from our unknown item

to a bank of signatures of items

that we would expect to find

or we have for expected munitions and other items.

Here on the left,

we have something that’s typical of a target of interest

or TOI.

It’s a 37-millimeter projectile.

And you can see there,

it’s got one strong primary polarizability

and two weaker and equal secondary

and tertiary polarizabilities.

This is typical of what we expect to see for munitions

because of their actual symmetry.

They’re mostly pipe type shapes.

Non targets of interest or none TOI, things like horseshoes,

scrap metal, the debris.

These typically have different polarizabilities,

they tend not to be equal.

They tend to sort of just be very irregular

because that’s what the shape of

most scrap pieces of metal are.

They are regular in shape.

And to give you an idea of,

you know, how well these kind of things work,

we can look here at a couple of different items.

Here we have a 37 and 75-millimeter.

They kind of have a different shape

but you can see clearly they have a different size,

see where they sort of would be coming in

and that Y axis intercept is located.

And this one always amazes me.

We can even detect and tell the presence of something

such as the driving band.

The driving band is usually a thin band of soft metal,

often copper, that is around the shell

that cause it to rifle or spin

as it travels through the barrel.

And whether that is located at the end of the round,

whether it’s lost during firing altogether

or it’s located in the middle round

causes slight changes in our polarizabilities.

And the fact that we can see that

I think is it’s pretty amazing and pretty cool stuff.

It does point out that we need to make sure that

our classification system and our methodology

can deal with subtypes, maybe damage to the item.

Also possibly just noise and inversion errors

as slight errors as we go along through our process.

So doing an Advanced Geophysical Classification survey

or AGC kind of comes down into two parts these days.

There is work being carried out

by the hardware manufacturer and others

for us to be able to detect and classify in a single pass.

But right now, we currently need to do things in two passes.

We do it as in a dynamic survey, kind of a mapping mode.

It’s kind of like mowing the grass

where we find all the possible locations

that we may have a source.

This can be done with conventional sensors,

things like the EM sensors that you’re familiar with,

magnetometers, you can use those,

but where appropriate the advanced EM sensors

do give you more accurate locations

and make the second part of the survey more efficient

because of these improved locations.

The second half of the survey is the

sort of the static survey or classification survey,

where we go and park our sensor at a flag location

and collect data to classify that source.

To give you some idea of production rates,

it’s often depending on the nature of the site,

how far the locations are apart.

We’ve found that people can collect roughly

three to 400 locations per day.

So looking at the dynamic survey, that mapping mode,

kind of breaks down into three easy steps.

We’re going to prepare the data, identify the sources

and they review those sources and create our list,

our flag list that we will use again in the static survey.

Some people like to call

the last two parts of this workflow,

the informed source selection or ISS,

because we’re using some idea

or knowledge of the sources that we’re looking for

to help us pick those targets.

In your UX-Analyze,

we have an easy workflow that will step you through this.

Here I’ve highlighted some of the key items

with the same cause as we just saw in the general workflow.

The items with the arrows on are the ones that

I would call them the must do’s.

That if you’re going to process data,

these are the things that we’re going to need to

step through.

So why don’t we take a moment here

and I’ll flip to Oasis Montaj

and we can take a look at

some of these parts of the workflow.

So here I have some data. I’ve imported it already.

And to save time,

I’ve gone through some of the processing steps,

because with the dynamic data

you do collect very large volumes of data

and it just takes sometimes,

a few moments for us to go through and do that processing.

The first step that you would do is do the data processing.

This is where we will filter and make sure

that any data that is outside

of our quality control specifications

is dummy down or removed from subsequent processing.

Like if the sample stations are too far apart because

the guys in the field went too fast

or there was some other problem with the sensor.

When this runs it will give you

sort of a plot similar to that.

And as we wrote down through our processing workflow,

the next thing you’ll need to do

is do some latency correction

for just timing aspects of how fast the system fires,

the GPS, all that kind of coming together.

Create the located database and then beta grid that up.

And I’ve got one of these here where I’ve prepared that

and shown it on here with our survey tracks.

One of the tips that I’d like to share with you,

often I like to see where was the data collected?

What you’re seeing there on that path there

is the original path of the cart of the sensor itself.

But remember it’s got in the case of this,

this is a two by two system, it’s got, as I was showing you,

those are the diagrams, those four receivers on those.

Well, where were those traveling?

And we can load and display those,

but whoa, that is just crazy.

I can’t see anything there.

One of my tips is set the transparency,

lower the transparency down of something of that

sort of other paths, where we can still see them.

We can still see where the original, the cart went

and then where those air receiver coils

traveled across our dataset.

And one of the things that

once we’ve created this amplitude grid,

one of the other steps that we like to look at

is something we call the coherence anomaly.

And this is where we look at a sample of the data

and see how well an item will model

under a window of that data.

And I’ll show you some examples of the threshold plots

in a moment.

The coherence anomaly map, let’s just move this over here.

I have gone and created one.

It makes it very easy to detect targets which you may miss

in the amplitude plot.

Now maybe we’d like to see our line paths on here as well.

And since I’ve already created them

over on my amplitude map,

it’s as easy as just I can drag and drop them onto this map

and we can see them there.

And they’ll come across

with their same transparencies and everything that we have

on the previous map.

And don’t forget, if you’re looking to look at multiple maps

and gee, wouldn’t it be nice if they were at the same place.

At the top of the map there is a series of buttons.

If I click on the second button from the left,

that will make all my maps zoom to the same area

that I see on the map that I am controlling it from.

For those, if you have run that UX-Analyze before,

any of you have noticed areas where you might see this

or some of these like broad high features

in your coherence anomaly data?

This is generally caused by over-correcting your data

or over leveling your data.

When you’re doing the leveling,

look at going in and adjusting parameters,

the leveling of filtering the data

to changing perhaps your amplitude threshold

or the width of the signal

that you’re using to do the filtering.

It says those types of areas where you see the broad,

in the broad high in the coherence anomaly,

or perhaps a broad low in the amplitude

is an indication that you’ve over leveled

or over filtered your data.

And after I went through and adjusted that,

we can see how I can make those a little bit more clearer.

Now I mentioned earlier that the coherence anomaly

allows us to see anomalies which we might not

be readily see in the amplitude.

Here I’ve just got to made a match snapshot,

you know, if you send maps to your friends

or to your manager or senior scientists to have a look at

and review?

So here, if I want to send this map on,

I’ll say, Dave, look, there’s a spot here

where I’ve got these low amplitude anomalies

and they’re only coherence elements.

What do you think?

He’s like, you know, rather and he’s got to look at this data

and go, well, where does he Darren want me to look?

He can just come unload that snapshot

and it will go to that area

and if he turns on the changing stance on all maps,

you can now quite easily,

we can go on and turn on our shadow cursor

and see which of those anomalies

we can see quite clearly on the coherence map,

but not so much in just the amplitude or response alone.

We would go on after we’ve, can pick our anomalies

from both the coherence and the amplitude picks

and using the thresholding tools,

we can sort of decide which threshold we used.

In this data set I used 0.5 and three

and I’ll show you in a moment how I picked those

when I flipped back to the slides.

And finally, you will invert the data

and generate the sources and then you need to be to filter

and look at those sources

and determine which ones are something

that you would like to go on and go on to

and collect a static survey data.

And I had a, so I just took slow open up my source database

where I’ve gone and done this.

In the source database.

So we go from survey data, we pick targets,

we then invert those targets to generate sources.

And then we might learn to look at,

and I’ll just overwrite the one I made earlier

of being able to filter out some of the sources

because some of them may be things which just,

there’s no way that they can be the

type of target of interest or UXO that we’re looking for.

And to make this easier or to help you with that,

when you do the filtering step,

we create a bunch of channels and we look for

how big was the source?

Did we get a good inversion result?

If we didn’t get a good inversion result

because of noise or something in the data,

then we can’t trust that result.

And we want to keep that on our list

for further investigation.

Some things we might say,

well, look, there’s no way it could have a size and decay

that that could be something that we could

possibly classify.

And these then are indicated in one of the channels here,

they’re sort of flagged off of that.

And there’s a nice, clear channel that tells you

or a column in the data as to why they were filtered out.

And you can see those various symbols represented on this

feature space of scatter plot.

The red lines represent my thresholds for my size and decay.

So you can see symbols out there turned off or gray,

whether it be green and turned on in here,

some of them were found just to simply have no anomaly

like we had no signal

in that when we do some of the inversions,

we will look for what we call three dipoles, three objects.

And one of the dipoles will have an object.

One or two of the others may not

if there’s only one physical object there.

And those will be filtered out and removed.

Some of the ones you see here with a little brown Xs

or orange Xs all over the site and some just on the edge,

these are the ones that due to some noise in the data

that we had a poor inversion result.

And we want to keep a new ongoing revisit in our static survey

to make sure that there is no targets of interest there

’cause remember, we’re dealing with UXOs here,

we want to be conservative.

So there’s a little bit of a high-level overview

and a few tips on the dynamic processing.

And just flipping back to our slides here,

we kind of walked through some of that workflow,

and I promise you to looking at the coherence threshold

or how we pick those thresholds.

So there’s a tool called

determined coherence anomaly threshold

that creates this plot for you.

And one of the questions I often get asked is,

well, how many samples should we use when we do this tool?

And you can see there on the right

I’ve just kind of highlighted where that sample box is.

We generally recommend that people use around

about 2,000 samples.

And the question is, well, why do I need to use 2,000?

I want to get it done faster. I want to use less.

This example where I use just 20 points,

where we go and find a background area.

So an area that we believe

because of its signal is free of metallic debris.

We synthetically insert a signal response into that

and invert it and see how well that inversion goes.

If it’s a very good match to something,

that gives us a strong coherence.

And we can look at how noisy that is

compared to when we just try to invert it

when nothing’s there.

With the object, gives us the black dots.

Without the synthetic object there,

it gives us the red dots.

And there’s the thresholds that I picked before

in our example.

And you can see there’s with 20 points.

If I ran it another time with this 20 points,

because we randomly picked the locations,

you get a slightly different result.

So maybe 20 is not good enough.

So use 200.

It’s better,

but I run it with another 200 points.

Things will slow the shift again.

Yes, I did run a couple of examples

to run several runs to cherry pick,

to give you some examples where you could clearly see

the shifts occurring.

But if I ran it with 2,000 points,

all of those variations that you see do get covered in,

it gives you a much more reliable set of curves

to bid a pick from, making an awesome,

makes them very easy to interpret and see.

If you’re wondering why the two levels

on the coherence plot,

depending on the nature of your munition,

we can get this sort of down curve into the y-axis.

And I went and picked a sort of more conservative point.

The purple line is the depth of investigation

that we’re being asked for on this project.

Is we’ve been asked to find things like

a 37-millimeter or something that can also be represented

by a medium ISO down to 30 centimeters.

ISO stands for Industry Standard Object.

And so that’s the level I want to pick for.

I want to be maybe a little conservative

and that’s why I’ve gone and lowered my threshold

down to 0.5.

Yes, could I go lower?

But that would be going sort of above and beyond

what we were asked for in our project scope.

So now we’ve chosen our thresholds.

There’s two places that we use the thresholds.

And my other little tip is when you take your values

that you’ve used in your target picking,

I ended up typing in three I guess or could have used 3.7.

I think I had 3.6 sort of in around there.

Those are the values that we picked targets with.

When we do the inversion because we have this full coverage

that we get with dynamic data.

If a source is found to be on the edge of the data chip,

that piece of data that we use to do the inversion with,

our results aren’t going to be as good or as reliable.

But because we have that full coverage,

we can reposition that chip.

During that repositioning

we want to look to see if we’ve got data there

and it’s got some signal to it.

Like if it’s not gone into total background.

And we recommend there for your thresholds

that you use roughly about 50%

of the original target picking threshold.

But be careful, don’t go down into the noise.

If your site tends to be a little bit noisier

in your original target picking threshold

is close to a noise threshold,

you might not be able to actually go that full 50%

and we’ll need to modify that value.

But on hand, generally,

you can use 50% of your target picking threshold

for your repositioning threshold

when we’re doing the inversions.

So that’s a bit of a walkthrough

and some tips around the dynamic workflow.

Remember if you have any questions,

do enter them into the questions box

and we’ll respond to you after the webinar.

Next here, I’ll take a look at the static survey

or classification workflow.

For this workflow there’s really just two steps.

Prepare the data and classifying rank.

There is this step in the middle there,

construct and validate a site library.

This is something you generally just do at the beginning

or end of your project to make sure

that the library that you’re using

is a complete library for your site.

And we’ll come and look at that

notion of a complete library for your site

a couple points through this part of the presentation.

So here’s our a static workflow.

Much like my title that seemed to be,

today seem to be very long

and might look to be very complicated.

But again, I can sort of highlight for you

with the same closes on a general workflow.

Those must do points are shown there with those arrows.

And it’s really, you know, import some data, level it.

If it’s the first time through

and you haven’t got a library,

you should validate your library

and then it’s just classifying rank.

But for most of your project import level classify.

And it’s pretty much that simple.

So let’s go over to Oasis and I’ll show you just

exactly how simple that is.

I’ve created a project,

I’ve already imported my background data

and then now we will go in and import my survey data.

I’ll just turn off the filter here.

In this folder

I’ve got all of my different types of survey files.

We have SAM or static anomaly measurements.

We have other QC ones which I’ll touch on

a little bit later as QC sense of function tests.

And then my background measurements as it says,

I’ve already done those.

The import with the HDF import,

you can just give it all of your data

and we’ll figure it and put the data

into the right databases for you based on its data type.

I’ve just gone and selected for this demonstration here,

just for files, static anomaly measurements over some items.

So we can import those in.

You’ll get two databases.

A data database and a target database

and there’s a similar ones in the dynamic workflow.

The data database, tongue twister it is,

contains the actual transient or response data.

The target database contains this a list,

in this case of all the measurements that you made

and various sort of other parameters

of the sensor that was UBIT size windows,

other premises that we use that describe the sensor

that we then use in the inversion.

So once I brought it in, I need to level it.

Leveling it is removing the geology or background

or drift or the sensor out of the readings.

So we’re here.

I just need to pick my survey database

and my background database.

We find those based on the codes

that are there in those names.

And then we will just a couple options that we need to pick.

Most people will pick time and location.

You going to pick the nearest background reading

that was taken in an area that was free of metal objects

and subtract that from our measurement.

And we want to use the one that’s the closest in space,

most likely the same geology and the closest in time

to remove any drift.

And that will give us a new level channel.

And now we are ready to do our classifying rank.

I’ve already got a library. It’s a good library.

I’ve done my validate library work.

So I can just hit classifying rank.

We give it our database to use.

Is asking us what channels you may be using

as a mass channel.

A mass channel is a way that you can turn off

individual targets or flag or measurements and not use them.

So if you just wanted to do a sample or redo something,

this is a way that you could toggle that.

Some parameters about the sensor,

but the library database that we’re going to match to,

I can go onto the tools tab and these are all the individual

sort of steps that this tool

of classifying rank will go through.

We’ll invert your sources, do the library match skill,

look to see if there’s any self matches or clusters,

identify those and ultimately do that set the thresholds

and do our classification and prioritization.

We can also create several plots.

I’ll just do the one plot at the end of this to speed it up

for our demonstration today.

So there we go.

That’s going to take about two minutes to run

and just kind of go through all of those steps for us

and generate plots for each of those targets that

we read in.

But while that’s running,

we can just take a look at

a data set that I have already brought in,

much more targets and have ran through,

sorry, this one.

And this is after I’d ran my validate library workflow.

Is where this one is at.

And some tips of things that are good to look out on here.

We will load a series of channels, there’s lots of data

in the database table that you can look at.

That’s just my progress bar.

But two that are good to add and show here

are the ones that look for clusters.

So I can hit down our list,

drop down and there’s are two clusters, channels.

And you can see over in my plot

where you see a whole bunch of gray lines

on this one that I happened to be sitting on in behind.

And that’s the notion of a cluster that they are the same,

has the same set of polarizabilities.

In the EM sense, they look alike.

And during the validation stage,

you want to look for these unknown items and find those.

And one of the easy ways to do that

is to come to our tool here.

That just keeps popping up

and just do show a symbol profile.

And in your profile window, we’ll give you,

we create an ID for each of the unknown clusters

and in this window now we can sort of, you know, see those.

And so we can see this like three lines here if you would,

that these guys are all the same as this

’cause this is the ID for this cluster.

So is a cluster called number two.

And there’s a bunch of things that look like each other

that are in this cluster.

And I can flick between those.

I’m just going to move that to my second screen

and it will generate and bring up the plots for those.

And we can see, well, oops, didn’t click on it quite there.

There we go.

Click into the database and you can see it of those.

So as a time saving tip,

when you’re doing your validation of your library,

you can bring this up

and just do a simple plot of the cluster IDs

and see quite easily then any of the things

of the eye of the clusters that we can explain

in that they don’t have a strong match

to something in our library.

Maybe you want to look at, well, what did they match to?

And you could bring up one of the other plots

which I have shown up here in the top right.

But I could show those what they match to

and what the match metric,

how well those curves match to each other.

We come up with a quantitative value there.

I could load those,

but one of the key powers of using the tables,

because we have a lot of information

is to use the database views.

And I have gone prior to this and saved myself a view

that I would like to have and I can get that.

And just by simply coming in and loading my view,

I can do that same step that I did

and showed you manually a moment ago,

loading those cluster channels,

but I can also then go and load a whole bunch of

whatever other channels I would like to look at.

Here look where we’ve matched all three curves,

one, one, one.

Where we’ve just matched two of the curves are primary

and the secondary curve.

And what did they all match to.

From a real power user point of view

for you guys there that are advanced users,

these database of view files are text files.

You can go in and edit them.

There’s the one that I just loaded up.

When we run the bundles,

we give you an automatically load up a database view.

If you don’t like or continually want to add things

to the database view that we have,

you can go and edit that file.

And add your own channels that you would like to see

loaded in there.

And then every time you run classifying rank

or the validate library,

that series of channels will be loaded.

So at this point a moment ago, when we saw there was

progress priors stop are example that we did there.

We brought in in here has completed.

It’s found that those items match to

and we’ve got some pretty good match metrics.

And if I click on the item,

we will automatically bring up

and show us one of the plots.

And we can see the polarizabilities on here.

We can see the polarizabilities

and how well that they’ve matched.

We can see where the sensor was parked and laid out.

They parked really good,

right over top of the potential flag location

of where the source is

and it was found to be at that location.

And will show you some other plots

which I’ll come to in a moment.

There’s our size and decay friend again,

and this other one called the decision plot.

And I can just kind of go through and look at those.

This one didn’t match as quite as well.

You can see the polarizability is from our data

and this one with a little bit noisier,

and we didn’t get still matched to

a one 20-millimeter projectile,

but just compared to one of those first ones I looked at,

they were just a little bit more noisier.

And so you’d see that was just as easy as import data,

level it and run classifying rank.

When we want to come and look at our results,

that was the one we were looking at before.

Here’s where I have gone and ran a larger sample.

And you can see there’s our size and decay plot

that gives us that,

looking at that feature of those two properties

and where things might group.

They’re colored based on red, we think it’s a TOI,

target of interest.

Green, it’s a below some threshold that we set

of how well things match to library items.

And that’s shown here in a little,

what we call a decision plot.

We can bring other plots up

or any other images that you may have on

that you want to see in your data.

We can create what these plots

that we call interactive image of yours.

This could be any image that you have.

Some people like to do this when they do their dig surveys,

that they will use photographs of the items.

And creating a view as simply as looking at the images

that you’ve created.

So this is my folder of some images that I,

polarization plots that I created earlier.

Seeing what the name is there and generally

we’ve got everything there with a prefix.

And then the name of the item is the last part of that.

And in this case that name is referring to

its initial acquisition ID.

So I’m going to go find that in the list here

and then browse into that folder.

And we’ll look in that folder and try to figure out

whether you’ve got a prefix as in my case, or,

you know, maybe you’ve done a suffix on there

for naming all the images and then whether the PNG, bitmap,

JPEG, whatever and then you can load it.

And then when I click on any one of our sources,

that image will load up along with my other image.

I can see here and we can look at it

and help you do the review.

Now, maybe you’re like me,

and you’ve got all these over and you’d be like,

gee, it would be nice if I could get them to arrange

the way that I would like them to arrange.

And for that, you could go,

we have a tool that will let you save a window layout.

Earlier I saved a window layout

and I had one from a self to get started,

but then I also had one for my classifying rank.

So I can load that one.

Oh, wait.

Sorry, I clicked on the wrong button there.

So is easy as that.

Now my window’s all arranged

in a way that I would like them to be

and I can easily move through

and they will update as I go forward and look at the plots.

We also have other tools.

I have a size and decay plot on that

main documentation plot that I made,

but maybe I would like to have an interactive version

of one of those so that when I click on it,

things will change and we can open up

and create one of those.

You can load overlays onto these plots

to help you find some of the items

that you were looking for.

And I’ve made an overlay earlier.

You can color them up based on one of the parameters.

If you do that based on the category,

we’ve got already a color code pattern in there

that will load.

So you can use that and this is always interactive

so that when I click on one of these items,

on the scatter plot, my database will go to it.

And as long with the other plots

that I am looking at and reviewing.

So there’s just a little bit of a walkthrough.

And I’m looking at some features with the static workflow.

And I’ll just kind of, you know, go back to my slides here.

And there was a couple of other,

you could say, frequently asked questions

I get from people, how to create a library.

Well, the answer’s real simple.

Collect some data over a target of interest.

With UX-Analyze we include an example library,

and this is based off a assertive ESTC project.

And I’ve included the link there

if you’d like to know about it.

But you can see some examples from that project where

they just took a sensor.

In this case, a MetalMapper,

put it on some plastic book stands

and then placed items underneath it and collected some data

that we then use as part of our libraries.

And as I touched on it in the demonstration.

Well, once these unknown items that are not on my site,

sorry, are not in my library, that are on my site.

Part of the validate library tool

looks for clusters of similar items

and then identifies the ones for you

which we don’t have an explanation for

that not in our library.

Then it’s up to you to go out,

collect some ground truth.

We identify which one is the most similar.

If you’re going to dig any of them up,

at least dig that one up.

Figure out what it is

and then there’s a tool where we can

add the item to our library as either a target of interest

or not a target of interest.

And then it will be used in the classifications


We have a number of quality control tests

because we want to make sure that the data

that we base our classification decisions on

is of the highest quality possible.

You don’t have time to go into those today,

but these have been developed to prevent common issues

and have all come from user feedback.

And there’s some,

it says those are the three main ones there.

How do we know this stuff works?

Oh, wait, sorry.


As you saw there,

the database tables can be easily exported out of

Oasis Montaj and then used in your reports.

All the plots that we generate are also saved as PNG files

so that you can then easily print them off,

include them in an appendix if you want to do that

and to share them among project stakeholders.

So how do we know that these things work?

Well we can look at something called

a receiver operating characteristic curves.

For a number of the demonstration projects of being done

while testing this types of technology.

They went and dug up all the items on the site

to see how successful the classification was.

So just there,

we have a sort of schematic of a rank dig list.

Red items are high confidence TOI,

yellow, most likely TOI and go on and dig them up

and then green high confidence none TOI.

Perfect classification, we would see a nice L

as shown on the left there or in the middle.

And if we just did a random guess 50, 50 guess,

you’d get something like a 45-degree line

on a one of these plots.

There is a tool in UX-Analyze.

If you want to do this to help calculate or create these

receiver operating characteristic or rock curves.

And from a couple of examples of the demonstration projects.

You can see, not quite at that sort of,

you know, perfect L-shape, but we’re pretty close

on some of them.

Some of these don’t go through zero by the way,

is because they asked for examples.

I hear they went and dug up some items to help them learn

or train the classification system.

So that’s how we know that we can reliably go and detect

and classify munitions or targets of interest

based on a geophysical response.

So in summary, today we’ve taken a look at

what is Advanced Geophysical Classification,

how we can use the EM response

to determine intrinsic property of the source,

the polarizabilities.

And those polarizabilities can be used for classification,

both whether it’s just a simple

matching to a physical property,

like size, shape, wall thickness.

Or actually using the signature mapping.

At the beginning there we saw how using

a method of classification

and being able to eliminate the clutter items.

We could reduce our excavation costs

and saving time and money on our projects.

We took a little bit of a look at dynamic

using dynamic advanced EM data

that gives you an improved target detection

and positioning over the conventional sensors

but you can use conventional sensors for the dynamic phase

is perfectly okay.

You’ll find you just have a few more re-shorts

when you’re doing your static classification survey.

And it’s possible with the static survey data

to reliably classify sources of geophysical anomalies

whether it’s a target of interest or not.

If you’d like to learn more about

Advanced Geophysical Classification,

we’ve got a couple online resources for you.

Going to their, there’s some papers and other presentations

that we’ve done and other examples.

If you’d like training, contact your local office,

so we can do remote training these days

or at some point set up a in-person training.

There is some material online

where you can go and also register for

some of our other training sessions.

Of course there is our support line

which you can read through a direct email

or through our website.

So hopefully you found this interesting and useful,

and I’d like to thank you for your time today.

And there’s my email again if you didn’t catch it earlier,

and thanks for your time

and hope you have a nice day.