DETECTING TRANSIENT WEATHER PHENOMENA ON MARS

An Overview of the Design and Calibration of Multiparametric Detection

Algorithms for the REMS/MSL Mission

J. Verdasca, J.-A. Manfredi and V. Peinado

Centro de Astrobiologia, Instituto Nacional de Técnica Aeroespacial

Ctra. de Torrejón a Ajalvir Km 4, Torrejón de Ardoz, Madrid, Spain

Keywords: Multiple-scale phenomena, change detection, autonomous decision, multiparametric algorithms.

Abstract: In its baseline operation mode, the REMS/MSL mission is not well suited to characterize transient weather

phenomena evolving on multiple scales. In this paper we argue that this limitation can be overcome by

enabling REMS to autonomously detect sudden and unexpected changes in the acquired data and

subsequently extend the time allocated for measurements. Detection is accomplished by means of

multiparametric Event-Finding Algorithms running on the REMS micro-controller. We describe the

principles underlying the design of such algorithms and the methods for their calibration.

1 DESCRIPTION OF REMS

The Rover Environmental Monitoring Station

(REMS) is an instrument suite to be carried aboard

the Mars Science Laboratory (MSL) rover. It is

being developed and integrated at the Centro de

Astrobiologia (CAB

/(CSIC-INTA)) as a contribution

of the Spanish Government. REMS comprises two

booms attached to the rover mast at an angle of 120

degrees (Figure 1).

Figure 1: REMS booms with sensors.

Each of the booms carries a wind sensor and an air

temperature sensor. An infrared sensor mounted on

the electronics casing of boom 1, facing the ground,

will measure the brightness temperature of the

martian surface. On boom 2 there is a humidity

sensor (the protruding cylindrical structure on the

figure) provided by the Finnish Meteorological

Institute (FMI). A pressure sensor, (not shown here)

also built by the FMI, is mounted below the rover

deck, being in contact with the atmosphere through a

small opening. Finally, an UV sensor (also not

shown), located on top of the rover deck, will

measure the irradiation intensity in six different

bands. REMS features a small embedded computer

system whose main task is to act as an Interface

Control Unit but is also capable of executing simple

programs.

2 THE REMS/MSL MISSION

2.1 Science Goals

The MSL mission will assess the present and past

habitability of Mars. REMS will contribute by

monitoring the physical processes taking place in the

layer of the atmosphere closest to the surface.

Scientist believe that if life exists on Mars it will

most likely be found in tiny pockets of liquid water

in porous rock beneath the surface. Local weather

patterns likely play a fundamental role in the ability

of sheltered environments to retain liquid subsurface

311

Verdasca J., -A. Manfredi J. and Peinado V. (2008).

DETECTING TRANSIENT WEATHER PHENOMENA ON MARS - An Overview of the Design and Calibration of Multiparametric Detection Algorithms

for the REMS/MSL Mission.

In Proceedings of the Fifth International Conference on Informatics in Control, Automation and Robotics - RA, pages 311-316

DOI: 10.5220/0001506703110316

 SciTePress

water. For instance, the regime of local winds may

impact the cycles of freezing and thawing of

permafrost soils inside craters and on the slopes of

mountains, cliffs or hills. Indeed, recent images

obtained from orbit show gullies and trenches that

seem to have been made by water outpouring at

certain localized spots and flowing down the slopes.

An alternative explanation, which rules out liquid

water, sustains that the gully-like features are caused

by small landslides triggered by wind. Hence, one of

the main goals of REMS is to shed light on the

regime of local winds to enhance our ability to

distinguish between features most likely due to

aeolian erosion from those which could possibly

have been carved by liquid water.

2.2 Limitations of REMS Baseline

The baseline currently being considered for REMS

Nominal Mode (NM) operation consists in 24

sessions per sol of 5 minutes each, comprising a data

acquisition cycle at a sampling rate of 1 Hz. Thus,

REMS ability to characterize short range multiple-

scale events is limited. In this position paper we

argue that this limitation can be overcome by

providing REMS with the capability to carry out

Opportunistic Science. We outline the strategy we

are pursuing to this aim, and lay down its scientific

basis. We show how, based on a sequence of

uploaded programs, the REMS computer can

autonomously decide to switch from NM to the so-

called Event Mode (EM) in case an anomalous

pattern is detected. When functioning in EM the

passage into Sleeping Mode is rescheduled to occur

after a configurable amount of time, during which

REMS will collect data at the nominal sampling rate.

3 OPPORTUNISTIC SCIENCE

ONBOARD MARS ROVERS

3.1 MER/OASIS vs. MSL/REMS

Opportunistic science onboard the Mars Exploration

Rover (MER) mission (Spirit and Opportunity) is

being implemented at NASA's Jet Propulsion

Laboratory (JPL) through a project known as the

Onboard Autonomous Science Investigation System,

(OASIS). This system identifies opportunistic

targets (mainly rock samples), collects data on those

targets, then analyzes them onboard to identify those

which merit further investigation. It also executes

onboard planning and scheduling software that can

dynamically modify the projected rover traverse

plan in order to accommodate the new science tasks.

Finally, the onboard computer evaluates if sufficient

resources are available to meet the additional science

requests (Castaño,

2007). Unlike OASIS, the REMS

Opportunistic Science computer programs assume a

purely passive role in planning and scheduling tasks.

The REMS computer executes a sequence of

instrument commands that the MSL main computer

has uploaded into its flash memory. These are

actions that REMS must execute at pre-scheduled

times without further intervention by the rover

computer – acquisition cycles, parameter

configuration, housekeeping, etc. The EFA's are part

of the set of programs uploaded into the REMS

memory and executed once every session. Their

only effect on the sequence of operations is to

(eventually) extend the period of data acquisition,

considering its previously allocated resources. This

framework is designed not to interfere with pre-

scheduled regular activities. The approach inevitably

limits the scope of REMS Opportunistic Science

activities but, on the other hand, improves its degree

of autonomy and risk avoidance level.

3.2 Targets for REMS Opportunistic

Science

The fast evolving phenomena that we wish to

sample normally involve multiple timescales which

cannot be fully accommodated within normal

session periodicity and/or duration. These are the

natural targets for Opportunistic Science actions.

They can be effectively classified into i) boundary

layer oscillations (BLO's), ii) dust devils and

plumes, iii) microfronts, iv) local dust storms, v)

orographic winds and vi) mountain waves. An

algorithm to detect BLO's must be capable of

recognizing fluctuations with timescales ranging

from a few minutes to about one hour. Dust devils

are much shorter events, generally not exceeding a

few minutes. They consist on rotating columns of

warm air made visible as sand and dust is sucked

into the central core. Usually, convective plumes are

even more ephemeral than dust-devils. Plumes are

characterized by straight-line winds converging to a

hot spot where the warmer air starts to rise. They do

not display the rotation of the horizontal wind

component that characterizes dust devils; rather, a

pronounced increase in the vertical wind component

is observed. Microfronts share the structure of their

synoptic counterparts but arise on a local scale

instead, taking a couple to a few hours to pass over a

given point. The evolution of global and regional

dust storms can be tracked from measurements in

REMS nominal mode. However, data will be way

ICINCO 2008 - International Conference on Informatics in Control, Automation and Robotics

312

too sparse to provide an accurate description of

short-scale, local events. The low atmospheric

density of the martian atmosphere and the intense

radiative processes drive robust orographic flows.

Furthermore, the patterns of large-scale circulation

are reinforced by the extreme regional asymmetries

of the Mars elevation field. Sometimes these

orographic flows appear associated with periodic

buoyancy perturbations which develop when stable

stratified air ascends a barrier – mountain waves.

Like on Earth, these waves occur on the scale of

individual mountains or chains with oscillation

periods ranging from a few minutes to one sol.

4 CHANGE DETECTION

ALGORITHM

4.1 Principles of EFA Design

Due to their multiple timescales, the above

phenomena would be missed or insufficiently

sampled should the unit operate always in NM. In

the case of phenomena lasting a few dozen minutes

to a few hours the measurements at high sampling

rate in sequences of short acquisition cycles will

mostly produce redundant data whereas the session

periodicity is much too low to permit a smooth

characterization. For the most brief phenomena,

such as plumes and dust devils, the session duration

would, in principle, be enough to fully capture the

evolution of one or more of these events, and the

sampling frequency adequate. However, because the

unit will be sleeping for more than 90 % of the time

the probability that such an event will fall within the

5 minute allocated slot is small; in most occasions

REMS will miss the phenomenon entirely. Indeed,

the experience of past missions, particularly

Pathfinder, suggests that relying on serendipitous

observations is a very poor strategy to detect dust

devils. Dedicated campaigns to search for occasional

phenomena are envisaged as part of an alternative

'human in the loop' strategy. Nevertheless, because

these are, in general, rare events with uncertain

periodicity they are difficult to target by

measurements to be carried out at pre-selected, fixed

times. One may, if the allocated measurement

window is too narrow, miss the phenomena

altogether. Alternatively, if it is too wide, valuable

resources will be wasted. Performing decision-

making onboard the rover allows for science goals to

be realized under circumstances where a control

loop from the ground is not possible because there is

not enough time for scientists to react to changes in

data patterns (by organizing dedicated campaigns,

for instance). However, the present framework is not

designed to replace entirely human decision making

but rather to optimize the combined action of human

experts and the REMS embedded system. REMS

planning and scheduling activities are performed on

the ground, by means of the Ground Data System

(GDS). The GDS provides all the computational

tools required to accomplish these tasks as well as

instrument health monitoring and data analysis. An

Opportunistic Science Software (OSS) package is

integrated in the GDS. Every day, the downloaded

REMS data product is processed at the GDS and the

part specific to the self-triggering procedure is fed

into the OSS. The complete statistics of event

detection is recorded for housekeeping, diagnosis

and debugging purposes. On a regular basis, analysis

of the results will permit fine tuning and in-flight

calibration of the EFA's. This procedure will be

supervised jointly by scientists and software

engineers present at the GDS. The standard

operation procedures involve generating activity

plans for a few sols determining, whenever possible,

if local meteorological events of interest are likely to

occur in that period. Having negotiated with the

MSL project management the resources for the

period being considered, the extra measurement time

available is distributed though a number of possible

EM windows. These and other relevant parameters

are updated at the next Schedule Table Script

(ST_Script) uplink. The REMS computer system is

then left to decide autonomously whether or not to

switch into EM according to the disposition of

allocated windows and the output of the EFA's.

4.2 Design Flexibility

The above framework requires an onboard software

design capable of offering full flexibility as the

process of calibration of the EFA's will continue on

Earth during the MSL cruise phase. Design

flexibility is also important so that diagnosis, fine-

tunning and debugging of the EFA's are possible

after operations begin on martian surface, with

minimal resort to software patching. It is also

expected to reduce substantially the risks to baseline

operation associated with the relative increase in

software complexity which the Opportunistic

Science capability inevitably carries. The current

design rationale is as follows: at the end of a

measuring session the REMS micro-controller will

run a prescribed sequence of EFA's. The order in

which the EFA's on the list are executed can be

modified at every ST_Script update. The first step in

DETECTING TRANSIENT WEATHER PHENOMENA ON MARS - An Overview of the Design and Calibration of

Multiparametric Detection Algorithms for the REMS/MSL Mission

313

any EFA is to process sensor raw data by applying

estimators to compute observables. The latter

emulate as close as possible the behavior of the

actual physical parameters (e.g. temperature,

pressure, wind magnitudes and direction, etc.).

Estimators are thus simple procedures that involve

functions and coefficients determined during the

sensor's calibration. The reliability of observables as

proxy values is assessed through comparison with

the output of the GDS post-processing software that

retrieves the 'true' physical parameters from raw

data. Each EFA is divided in two parts running

sequentially: detection and validation. Only

validated detections will cause REMS to self-trigger

into EM. The detection part of the EFA reads the

relevant observables and, through a sequence of

logical operations decides whether or not a detection

has taken place. Then it assigns an initial confidence

level to that detection, called the Baseline

Confidence Level (BCL) to be compared with the

uploaded Threshold Confidence Level (TCL); if

BCL > TCL, the detection is validated: the EFA will

immediately stop running, all subsequent EFA's on

the list will be overruled and the unit will shift into

EM at once. If at least one detection has occurred

but nevertheless had been assigned a BCL value

below threshold, then the EFA will compute a

number of flags and/or consult a table of previously

set flags. The final result of these computations is an

integer value, called the Increase in Confidence

Level (ICL), incremental to the BCL. Both intervene

to calculate the final Detection Confidence Level

(DCL). If DCL > TCL, then EM is triggered. Flags

are encoded signals either internally generated by

the EFA acting on observables or historic records, or

engendered at the GDS. In the later case, the flags

are generated by the OSS package and uploaded as

parameters in the ST_Script. Flags have limited

resolution (small integers and binary numbers) and

intervene only at the validation stage. For instance,

the orography flag may take up to 16 integer values

representing a type of orographic feature – crater

rim, mountain ridge, rock, etc. – a scale of elevation

with respect to rover position and the feature's size.

Internal generation of flags, within the EFA, will

sometimes involve elementary statistical analysis

while in other cases it implicates logical operations

and/or Lookup Table (LUT) consultation. For

instance, in the EFA for microfronts described in the

next section, the PLO flag – for 'Pressure Low' –

indicates, based on the history of measurements

performed in past sessions, whether a sustained

pressure decrease in the hours preceding the passage

of a cold front has occurred. When no detection has

taken place, or achieved the TCL required for

validation, the system proceeds to execute the next

EFA on the list. If, once all the EFA's on the list

have been executed, no detection occurred or was

validated, the unit will continue its scheduled

operations normally, going into sleeping mode at the

pre-scheduled time. REMS will keep in memory the

session averages of the observables as well as a

history of the parameters involved in the EFA's,

such as confidence levels, flags, etc. Normally, data

for a complete sol should be kept but this may vary

depending on total memory load.

4.3 Two-tier Example Algorithm

We provide an example of EFA structure in Figure 2

using a simplified flow-chart. The example EFA

considered here targets microfronts. It is a Multiple-

Session Detection Algorithm (MSDA), i.e entails

accessing data from more than one session. Note

also that it requires detrending, i.e the deviations

with respect to the foreseeable trend have to be

calculated in the beginning of the algorithm. A first

level detection, L1, results from a qualitative

divergence with respect to the trend, when

temperature rises in late afternoon or drops before its

normal apex, for instance. The persistence of this

inversion for a number of consecutive sessions is

interpreted as a sign of confidence build up, and thus

the Baseline Confidence Level (BCL) is set

accordingly. A second detection level, L2, is based

on quantitative divergence only. Even in the case

when the sign of temperature variation is the one

expected, there still may be a discrepancy in its

quantitative rate-of-change (ROC). For instance, the

ROC value may exceed some threshold d above the

expected temperature progression. As this condition

is obviously weaker than trend inversion it shall be

assigned a lower BCL. If the relevant BCL is lower

than the threshold, then the EFA will proceed to the

compute flags – in this case the PLO flag mentioned

above and also two other flags signaling,

respectively, an increase in wind magnitude (WIM)

and the persistence in direction (DIR) of the wind

vector. Depending on their values, the DCL's may or

may not be increased, as previously described.

5 FUNCTIONALITY TESTS AND

CALIBRATION

5.1 Overview of the Procedure

Functionality tests and calibration of the EFA's will

be carried out on a desktop computer emulation

using selected signatures of the events as input.

These test signatures can be created from: i) real

ICINCO 2008 - International Conference on Informatics in Control, Automation and Robotics

314

Figure 2: An example EFA flow-chart.

Martian weather data ii) output of mesoscale models

ii) meteorological data from martian environmental

analogues iii) engineered representative datasets and

iv) field tests. Upon application of the fine-tunning

procedure the percentage of false positives and false

negatives is evaluated and once their numbers lie

within acceptable levels the EFA is considered

calibrated.

5.2 Martian Weather Data

Up until now only the Viking and Pathfinder

missions have carried out successful meteorological

experiments on the surface of Mars. The Phoenix

lander - currently in its cruise phase and due to land

on May 25, 2008 – carries temperature and pressure

sensors together with wind speed and direction

indicators. Currently we are working with data from

Viking and Pathfinder and, if Phoenix is successful,

we shall use its data as well. Each of those datasets

has its merits and faults: the Viking data have poor

resolution but is extensive; Pathfinder data cover a

much shorter period but display higher resolution. In

the later case, in addition to signatures of dust devils

registered by the sensors, there are, in some cases,

simultaneous confirmation from images. We shall

apply the EFA's to selected parts of the time series

chosen because they reveal the characteristics of a

given event, with particular focus on the patterns

already presented in the scientific literature as

symptomatic of the phenomena under scrutiny. After

a number of refinement stages we expect to reach

the minimal algorithmic complexity required to

detect the relevant features.

5.3 Martian Environmental Analogues

and Field Tests

Because weather data from Mars is scarce we will

have to use data from martian environmental

analogues and field tests on Earth. Overall, the best

candidates for the role of environmental analogues

are the dry valleys of Antarctica, because of the

extremely cold and dry air and the fact that a fairly

predictable regime of robust slope winds exists.

When searching for dust devils and plumes the best

places to look for on Earth are deserts. These

phenomena have been extensively studied in the

Arizona desert (Renno, 2004). It is an unfortunate

yet unavoidable fact that the optimal locations to

conduct field tests are harsh environments in remote

areas of difficult (and thus expensive) access. This

fact poses a constrain in our current development of

a priority list for site selection for field tests. The

sites being evaluated include Deception Island –

home to a Spanish Antarctic base, where members of

the REMS team regularly conduct research – as well

as sites in the Atlas mountains and the Atacama

desert. Nevertheless, locations in close reach of the

CAB facilities will be favored, at least in the

preparatory phases of the field test campaign. In this

vein, we are currently assessing the reliability of

certain arid and semi-arid regions in the Iberian

peninsula as partial analogues. A first set of tests

will be conducted with a commercial weather

station. In a later stage, we shall use a REMS

model, complete with rover mast and deck dummies,

to carry out end-to-end performance and reliability

tests of the EFA's under the geometrical constraints

imposed by integration with the rover.

DETECTING TRANSIENT WEATHER PHENOMENA ON MARS - An Overview of the Design and Calibration of

Multiparametric Detection Algorithms for the REMS/MSL Mission

315

5.4 Engineered Artificial Datasets

In order to overcome the difficulties of using

terrestrial data in calibration, due to dissimilarities

between the atmospheres of the two planets, we shall

work also with engineered data. Based on the

analysis of field data and sensor calibration data

obtained under martian conditions, we shall create

artificial datasets to mimic the signatures expected

from the phenomena. Integrating sensor calibration

data with theoretical models and/or computer

simulations, notably mesoscale models of the

atmosphere, we can also create prototype signatures

that minimize the constraints set by wind tunnels

and martian atmosphere simulation chambers.

6 ANTICIPATION

We have outlined a strategy to enable autonomous

decision-making aboard a mobile robotic surface

probe, in order to optimize its science return.

Provided with this capability, REMS will be able to

detect sudden and unexpected changes in

environmental parameters adjusting its operation

mode in response, without the need for a control

loop from Earth. As planetary exploration stretches

out further away from the Earth, communication

delays and interruptions will make remote control

increasingly harder. Though in the case of flyby and

orbital missions these delays can be accommodated

without much hindrance the situation is very

different for robotic probes evolving in a changing

environment. For a vehicle exploring Titan's surface

or Europa's putative ocean, onboard intelligent

systems will be essential for virtually all craft

operations including guidance and steering, obstacle

avoidance, selection of scientific targets, data

downlink management, etc. Sending probes to land

on these far away bodies is not envisaged in the near

future. Until then, the surface of Mars will the

cutting-edge operation scenario and the most

demanding testing ground for the implementation of

autonomous decision-making in the context of

planetary exploration by robots. Inspired by the

OASIS concept (http://ml.jpl.nasa.gov/oasis/), we

decided to use REMS for an experimentation in

Opportunistic Science. However, because the

processing power of the REMS computer system is

low, (even when compared to common household

appliances), we are forced to use exceedingly simple

programs whose ability to identify changes in

environmental conditions cannot be guaranteed a

priori. But even if some of the concepts being

developed do no mature fast enough to be

implemented in REMS/MSL, they still may find

application in future missions, to Mars and beyond.

Because operational autonomy is intrinsic to the

REMS design, the software under development

needs not to be integrated with the spacecraft's flight

software, and thus the employment of an upgraded

version in future missions should be straightforward.

This tempts us to anticipate a bold strategy. Indeed,

one can envisage a scenario in which future landers

and rovers being sent to Mars would each carry a

station similar to REMS, provided with autonomous

decision-making abilities. Thus, even after every

other component of the mission has reached its end-

of-life the spacecraft would not turn into waste but

instead become a node in an ad-hoc network of

weather stations. Of course, each future autonomous

station would need dedicated power and data

downlink systems; and would have to operate at a

low sampling rate and use recurrent transfer

windows in order to lower its power consumption

and step-up the probability that the data streams be

received unbroken by a (future) small network of

low-orbit communication relay satellites. This

network could be considered either as a precursor

mission or a scalable inexpensive alternative to the

MetNet mission (

http://www.ava.fmi.fi/metnet-portal/),

currently being proposed, and also NetLander

(

http://smsc.cnes.fr/NETLANDER/), abandoned in 2003.

REFERENCES

R. Castaño et al., Journal of Field Robotics, 24 (5), 379-

397, 2007.

N. Renno et al, Journal of Geophysical Research, 109,

E07001, 2004.

ICINCO 2008 - International Conference on Informatics in Control, Automation and Robotics

316