Intensive Monitoring Survey (IMS) Strategies

• IMS is a high-cadence, domain-specific monitoring approach that combines repeated, dense temporal sampling with tailored inference methods to capture critical events near detection limits.
• In astronomy, IMS implementations like the 7DS and IMSNG leverage daily to hourly cadences and deep stacking to enhance spectral mapping and early transient detections such as supernova shock breakouts.
• Across fields, IMS frameworks integrate multimodal data—from optical and RF to blurred video and biosignals—while addressing domain-specific challenges like artifact reduction, privacy, and sampling bias.

to=arxiv_search েউjson {"query":"(Ko et al., 26 Sep 2025) Photometric Redshift Forecast for 7-Dimensional Sky Survey (Emberger et al., 2023, Im et al., 2019, Nguyen et al., 2023, Kitov, 2 Jun 2026, Clark et al., 2015)","max_results":6} Intensive Monitoring Survey (IMS) is a domain-dependent designation for high-cadence, repeatedly sampled monitoring programs that combine dense temporal coverage with downstream inference. In the current literature, the term refers most directly to the daily-cadence medium-band component of the 7-Dimensional Sky Survey, to the Intensive Monitoring Survey of Nearby Galaxies, and to survey- or framework-level constructs in clinical, indoor, and demographic monitoring; the same acronym also appears in seismology as the CTBTO International Monitoring System, which is a distinct usage rather than an Intensive Monitoring Survey (Ko et al., 26 Sep 2025, Im et al., 2019, Emberger et al., 2023, Nguyen et al., 2023, Clark et al., 2015, Kitov, 2 Jun 2026).

1. IMS in the 7-Dimensional Sky Survey

Within the 7-Dimensional Sky Survey (7DS), the Intensive Monitoring Survey is the daily-cadence spectral mapping component designed both to track short-term spectral variability, such as in AGN, and to deliver deep stacked medium-band photo-spectra for galaxy evolution and large-scale structure studies. Its field is a $\sim 20\ \mathrm{deg}^2$ area near the South Ecliptic Pole, observed daily over five years. The survey uses 40 medium-band filters spanning $\sim 400$–$900\ \mathrm{nm}$, each with $\mathrm{FWHM}=25\ \mathrm{nm}$, yielding spectral resolution $R \approx 50$. The 7DT telescope array observes at different wavelengths nearly simultaneously, which minimizes color errors from variability and enables time-series photo-spectra (Ko et al., 26 Sep 2025).

The mock-catalog construction and redshift simulations adopt EL-COSMOS model SEDs derived from COSMOS2015 over $0 \lesssim z \lesssim 2.5$, including stellar continua, nebular continuum, and the emission lines $[\mathrm{O\,II}]\ 3728\ \AA$, $\mathrm{H}\beta\ 4863\ \AA$, $[\mathrm{O\,III}]\ 4959/5007\ \AA$, and $\mathrm{H}\alpha\ 6565\ \AA$. Magnitudes are defined as $\sim 400$0, so $\sim 400$1 denotes the magnitude in the $\sim 400$2 filter. The five-year IMS exposure is scaled to an on-source total of $\sim 400$3 seconds across the 40 bands, assuming observing efficiency $\sim 400$4 and the survey cadence. Signal-to-noise is modeled as

$\sim 400$5

with the $\sim 400$6 terms representing source photons, sky background, detector dark current, and readout noise over the exposure time.

The stacked depth is central to the survey’s performance. Single-epoch IMS depth matches the Reference Imaging Survey, but the five-year stacked IMS depth reaches $\sim 400$7 at $\sim 400$8 for point-source detection, and the reported $\sim 400$9 depth at $900\ \mathrm{nm}$0 is $900\ \mathrm{nm}$1, compared with $900\ \mathrm{nm}$2 for WTS Y5 and $900\ \mathrm{nm}$3 for RIS. This makes IMS Y5 $900\ \mathrm{nm}$4 mag deeper than WTS Y5 and $900\ \mathrm{nm}$5 mag deeper than RIS at $900\ \mathrm{nm}$6 by stacking.

Photometric redshifts are computed with EAZY using the eazy_v1.3 PEGASE-based template set with emission lines added following Ilbert et al. (2009). The main estimator is $900\ \mathrm{nm}$7, defined by minimum $900\ \mathrm{nm}$8 with nonnegative linear combinations of templates and no prior. For IMS, the reported results are template-fitting only, without magnitude priors. The residual and error metrics are

$900\ \mathrm{nm}$9

$\mathrm{FWHM}=25\ \mathrm{nm}$0

$\mathrm{FWHM}=25\ \mathrm{nm}$1

and

$\mathrm{FWHM}=25\ \mathrm{nm}$2

$\mathrm{FWHM}=25\ \mathrm{nm}$3 bin	IMS Y5 metrics $\mathrm{FWHM}=25\ \mathrm{nm}$4
$\mathrm{FWHM}=25\ \mathrm{nm}$5	$\mathrm{FWHM}=25\ \mathrm{nm}$6
$\mathrm{FWHM}=25\ \mathrm{nm}$7	$\mathrm{FWHM}=25\ \mathrm{nm}$8
$\mathrm{FWHM}=25\ \mathrm{nm}$9	$R \approx 50$0
$R \approx 50$1	$R \approx 50$2

These results are substantially better than RIS at similar magnitudes and better than WTS Y5 in the faintest bin, where WTS Y5 is reported to have $R \approx 50$3 and $R \approx 50$4. The paper attributes this to deeper stacked $R \approx 50$5, which reduces failures close to the survey limit. The physical basis is the ability of $R \approx 50$6 medium bands to localize the $R \approx 50$7 and Balmer breaks and to isolate $R \approx 50$8, $R \approx 50$9, $0 \lesssim z \lesssim 2.5$0, and $0 \lesssim z \lesssim 2.5$1 within $0 \lesssim z \lesssim 2.5$2 passbands. For $0 \lesssim z \lesssim 2.5$3, where the $0 \lesssim z \lesssim 2.5$4 break and $0 \lesssim z \lesssim 2.5$5 traverse the optical, deeper stacking improves the robustness of both continuum and line measurements.

A related diagnostic is the color-excess method, exemplified by $0 \lesssim z \lesssim 2.5$6 as a function of redshift. Positive $0 \lesssim z \lesssim 2.5$7 tends to indicate that a strong emission line falls in the $0 \lesssim z \lesssim 2.5$8 band, while negative values align with spectral breaks. In the mock analysis, restricting to $0 \lesssim z \lesssim 2.5$9 and $[\mathrm{O\,II}]\ 3728\ \AA$0, a selection of strong color-excess galaxies with $[\mathrm{O\,II}]\ 3728\ \AA$1 produced 54 candidates with no catastrophic failures, whereas a control sample with $[\mathrm{O\,II}]\ 3728\ \AA$2 yielded 1,740 galaxies with 187 catastrophic failures. The paper also quantifies near-IR synergy only for WTS, not IMS; a plausible implication is that IMS would benefit most in the faintest and higher-redshift regimes from VIKING or SPHEREx, because the demonstrated WTS gains arise from breaking low-$[\mathrm{O\,II}]\ 3728\ \AA$3/high-$[\mathrm{O\,II}]\ 3728\ \AA$4 color degeneracies rather than from changes to the optical medium-band data themselves.

2. The Intensive Monitoring Survey of Nearby Galaxies

The Intensive Monitoring Survey of Nearby Galaxies (IMSNG) is a high-cadence, multi-site program designed to detect the earliest optical light from supernovae in nearby galaxies with high supernova probabilities. Its scientific objective is to constrain explosion mechanisms by inferring progenitor-system sizes from shock-heated emission lasting less than a few days after explosion. The survey monitors 60 galaxies selected to be near-ultraviolet bright, $[\mathrm{O\,II}]\ 3728\ \AA$5, within $[\mathrm{O\,II}]\ 3728\ \AA$6, and at Galactic latitude $[\mathrm{O\,II}]\ 3728\ \AA$7, with two low-latitude exceptions. The monitoring network uses $[\mathrm{O\,II}]\ 3728\ \AA$8-m to $[\mathrm{O\,II}]\ 3728\ \AA$9-m class telescopes distributed across Korea, Uzbekistan, Australia, and the United States, achieving cadences of hours at depths around $\mathrm{H}\beta\ 4863\ \AA$0 (Im et al., 2019).

The target selection is explicitly yield-oriented. The adopted NUV threshold corresponds roughly to $\mathrm{H}\beta\ 4863\ \AA$1, preferentially selecting actively star-forming, comparatively low-extinction systems. The survey estimates a supernova rate of $\mathrm{H}\beta\ 4863\ \AA$2 per galaxy, about $\mathrm{H}\beta\ 4863\ \AA$3 the canonical average of $\mathrm{H}\beta\ 4863\ \AA$4. With 60 galaxies, the expected yield is $\mathrm{H}\beta\ 4863\ \AA$5, with early light-curve coverage to $\mathrm{H}\beta\ 4863\ \AA$6. The paper reports that over five years 18 supernovae occurred in the target fields, 16 of them in IMSNG galaxies, confirming the estimated rate: the realized frequency is $\mathrm{H}\beta\ 4863\ \AA$7, or $\mathrm{H}\beta\ 4863\ \AA$8 per galaxy, consistent within Poisson errors with the prediction.

The observational logic is set by the physics of shock breakout and post-shock cooling. The cited theoretical framework links the characteristic timescale to progenitor radius through

$\mathrm{H}\beta\ 4863\ \AA$9

with cooling-envelope luminosity and temperature scaling schematically as

$[\mathrm{O\,III}]\ 4959/5007\ \AA$0

For SNe Ia, companion-interaction emission is anisotropic and scales approximately linearly with companion radius for favorable viewing angles. Operationally, IMSNG uses the detectability relation $[\mathrm{O\,III}]\ 4959/5007\ \AA$1. At $[\mathrm{O\,III}]\ 4959/5007\ \AA$2, the survey is designed to detect shock emission from $[\mathrm{O\,III}]\ 4959/5007\ \AA$3 progenitors under optimal viewing and timing; at $[\mathrm{O\,III}]\ 4959/5007\ \AA$4, it reaches $[\mathrm{O\,III}]\ 4959/5007\ \AA$5 under favorable conditions. The abstract summarizes the program as enabling detection of shock-heated emission from a progenitor star with a radius as small as $[\mathrm{O\,III}]\ 4959/5007\ \AA$6.

Cadence compression is achieved by longitude distribution. The nominal cadence is about one day per galaxy, but combined observations across Korea, Uzbekistan, and the United States provide $[\mathrm{O\,III}]\ 4959/5007\ \AA$7-hour coverage, and some equatorial targets observed from Korea and Australia reach $[\mathrm{O\,III}]\ 4959/5007\ \AA$8-hour cadence. Routine monitoring is primarily in $[\mathrm{O\,III}]\ 4959/5007\ \AA$9 or $\mathrm{H}\alpha\ 6565\ \AA$0, with exposures of 1–5 minutes; once a supernova is identified, multi-band sequences such as BVRI or griz are added to constrain color and temperature evolution.

The survey has also produced case studies and ancillary science. SN 2015F in NGC 2442 yielded very early light curves and a companion-radius constraint of $\mathrm{H}\alpha\ 6565\ \AA$1. SN 2017gax in NGC 1672 was captured before and after explosion, with the first IMSNG detection preceding the discovery image reported by another group. The accumulated imaging supports additional work on luminous red novae, luminous blue variable eruptions, AGN variability, variable stars, asteroids, and faint low-surface-brightness structures revealed by stacks reaching $\mathrm{H}\alpha\ 6565\ \AA$2. The program therefore couples a narrowly defined early-supernova objective to a broader time-domain and deep-imaging archive.

3. Privacy-preserving intensive monitoring in critical care

In critical-care monitoring, the supplied literature uses IMS as a framework for integrating biosignals with privacy-preserving video context. The underlying motivation is that high-resolution biosignals such as arterial pressure, intracranial pressure, oxygenation, ECG, and EEG are vulnerable to artifacts caused by patient motion and staff interventions, and that signal-only analysis often lacks the contextual information needed to distinguish physiological events from acquisition artifacts. The cited ICU study addresses this gap under two operational constraints: only severely blurred video can be stored, and computation must run on hospital hardware rather than in the cloud (Emberger et al., 2023).

The video pipeline is designed around standard object detection rather than custom video architectures. Frames are blurred using FFmpeg box blur with boxblur=6:1, and the method repurposes the three RGB channels to encode temporal information while remaining compatible with off-the-shelf detectors such as YOLOv5s. With current frame $\mathrm{H}\alpha\ 6565\ \AA$3, grayscale conversion $\mathrm{H}\alpha\ 6565\ \AA$4, motion indicator $\mathrm{H}\alpha\ 6565\ \AA$5, and a previous-frame bounding-box bitmap $\mathrm{H}\alpha\ 6565\ \AA$6, the encoded input is

$\mathrm{H}\alpha\ 6565\ \AA$7

The red channel preserves coarse shape information, the green channel emphasizes large pixel changes relative to the previous frame, and the blue channel encodes previous bounding boxes. For the blue channel, pixel values are set to 32 inside the union of previous boxes and 0 otherwise. To reduce over-reliance on prior detections, only a randomly selected half of samples include non-zero $\mathrm{H}\alpha\ 6565\ \AA$8; among those, 20% discard $\mathrm{H}\alpha\ 6565\ \AA$9 entirely, and 60% of included boxes are jittered by up to 10 pixels.

The operational setting is a 12-bed neurocritical care unit at University Hospital Zurich. Continuous bedside recordings at $\sim 400$00 and 25 fps were collected by an ICU Cockpit infrastructure; only blurred streams were stored. Motion-rich clips were identified by pixel-wise frame differences, padded by 10 seconds on both sides, and 30,748 clips were accumulated. Of these, 196 clips were manually labeled across the classes patient, bed, staff, and devices, with rotated bounding boxes permitted. The train/validation/test split was 70%/15%/15%. Both baseline and proposed models used default YOLOv5 hyperparameters, with learning rate 0.01 and momentum 0.937 after a 3-epoch warm-up, pre-warm-up learning rate 0.1 and momentum 0.8, and weight decay 0.0005. Training stopped after 100 epochs without improvement, up to a maximum of 300 epochs.

Performance is reported with $\sim 400$01, using

$\sim 400$02

The proposed temporal-channel method achieved $\sim 400$03, versus $\sim 400$04 for the baseline YOLOv5 model, a $\sim 400$05 absolute improvement. Per-class $\sim 400$06 was 99.5% versus 98.0% for bed, 58.1% versus 58.1% for staff, 98.4% versus 97.6% for devices, and 99.4% versus 95.3% for patient. Training also converged much faster: 10 epochs for the proposed method versus 119 for the baseline. Staff remained the most difficult class under severe blur, especially with occlusions, coverings, and low-light conditions.

The clinical significance lies in artifact attribution rather than generic object detection. The paper explicitly frames staff presence, patient posture or coverage, and device context as information that can be aligned with biosignal anomalies to reduce false alarms and support decision support systems. The method is therefore not a general theory of intensive monitoring, but a deployable privacy-preserving component for multimodal IMS-style monitoring in hospitals.

4. IMS as a multimodal indoor human monitoring architecture

A broader systems view appears in the survey of non-contact multimodal indoor human monitoring systems, where an IMS is characterized as an integrated sensing and machine-learning stack using heterogeneous modalities, primarily cameras and radio-frequency devices, with optional IMUs, pressure plates, microphones, and ambient sensors. The emphasis is on continuous, robust, privacy-conscious monitoring in homes and care facilities, particularly for elderly care. The survey organizes this IMS concept by task—activity recognition, vital-sign measurement, user identification, localization and tracking, and 3D depth or scene modeling—and by fusion paradigm: combination, transformation, and collaboration (Nguyen et al., 2023).

The sensor taxonomy is explicit. Cameras contribute RGB streams, depth maps, skeletons, silhouettes, and thermal imagery; common pipelines include YOLO-based detection, pose estimation with systems such as AlphaPose or OpenPose, tracklet formation, and silhouette features. Radio devices contribute WiFi CSI or RSSI, BLE AoA/AoD and RSSI, UWB time-of-arrival and ranging, FMCW or mmWave point clouds, Doppler or micro-Doppler signatures, and spectrograms. IMUs contribute tri-axial acceleration and gyroscope trajectories, often embedded in smartphones or wearables. The survey emphasizes the complementarity of modalities: visual streams offer high semantic richness but are lighting- and occlusion-sensitive and privacy-sensitive; RF sensing is robust to lighting and non-line-of-sight conditions but lacks texture and often requires substantial postprocessing.

The fusion formalism includes early, intermediate, and late fusion, as well as cross-modal transformations and co-learning. The survey gives transformer-style notation for cross-modal attention: $\sim 400$07

$\sim 400$08

and cross-modal information exchange through

$\sim 400$09

Spatial localization and activity spaces are represented as feature vectors $\sim 400$10, locations $\sim 400$11, and trajectories $\sim 400$12. The survey also notes the use of STFT spectrograms in several vital-sign systems, although it does not reproduce the explicit STFT formula.

Reported performance spans multiple tasks. Gait identification with mmWave radar plus RGB reached up to 95.4% accuracy through space-time feature concatenation; WiFi-plus-video gait identification achieved 94.2% accuracy with a two-stream LRCN plus LSTM fusion model; RFID plus Kinect ID assignment reached 96.6% accuracy within 4 seconds. For localization and tracking, camera plus RSS plus IMU systems achieved 0.7 m accuracy, UWB plus monocular camera reached $\sim 400$13, BLE 5.1 plus visual point cloud plus IMU attained median position error 8.4 cm and angular error $\sim 400$14, and radar-vision attention-based fusion reported $\sim 400$15 in distinguishing pedestrians from billboards. The survey’s characterization of an IMS is therefore strongly multimodal and task-complete rather than tied to a single sensing modality or estimator.

Datasets and deployment guidance reinforce that characterization. The survey catalogs multimodal benchmarks such as Berkeley MHAD, UTD-MHAD, UR Fall Detection, Opportunity++, OPERAnet, and radar-plus-RGB-D systems. It also recommends room-dependent modality allocation, such as vision plus RF in common spaces and RF-only sensing in bedrooms, along with synchronized acquisition, edge-aware computation, and privacy-aware operation in which RF can dominate routine sensing and cameras are activated or reviewed on demand. This suggests that, in this literature, IMS names a design envelope for multimodal non-contact monitoring rather than a specific instrument or survey field.

5. IMS as an HDSS-centered mortality monitoring design

A public-health formulation of IMS appears in the Hyak framework for mortality monitoring. Here an Intensive Monitoring Survey is built around a Health and Demographic Surveillance System (HDSS) core with “frequent, intense, linked, prospective” follow-up and a surrounding informed sample survey designed to capture as many death events as possible. Hyak has three named components: data amalgamation, cause of death, and socioeconomic status. Data amalgamation combines HDSS surveillance with survey sampling outside the HDSS; verbal autopsy is used to estimate cause-of-death distributions; and SES is measured both descriptively and as model covariates (Clark et al., 2015).

The sampling design is model-based. Using a historical HDSS cohort, mortality risk is fit for village $\sim 400$16 and stratum $\sim 400$17 by

$\sim 400$18

with four sex-age strata: young girls $\sim 400$19, young boys $\sim 400$20, older girls $\sim 400$21, and older boys $\sim 400$22. Fitted probabilities $\sim 400$23 are aggregated to village-level risk,

$\sim 400$24

and the total sample size $\sim 400$25 is allocated across villages in proportion to predicted deaths,

$\sim 400$26

An alternative optimum allocation is the Neyman rule,

$\sim 400$27

Estimation of total deaths uses

$\sim 400$28

where $\sim 400$29 is the observed number of sampled deaths and $\sim 400$30, $\sim 400$31 are the population and sample sizes for village $\sim 400$32, stratum $\sim 400$33. The framework compares progressively richer models, culminating in a spatial covariate logistic mixed model,

$\sim 400$34

with village random effect $\sim 400$35, household random effect $\sim 400$36, and spatial effect $\sim 400$37 following an intrinsic conditional autoregressive prior,

$\sim 400$38

Bayesian estimation is performed with INLA, using flat priors on $\sim 400$39 and $\sim 400$40 and Gamma priors $\sim 400$41 on variance components in the simulations.

The proof-of-concept simulation is based on the Agincourt HDSS context in rural South Africa, using 20 villages with child populations $\sim 400$42, 50% boys, 50% girls, and age composition 20% in $\sim 400$43 and 80% in $\sim 400$44. Sampling strategies include two-stage cluster sampling, stratified equal allocation, Hyak informed sampling, and optimum allocation. Total sample sizes are $\sim 400$45, with $\sim 400$46 replicated draws per design. Accuracy is evaluated by MSE,

$\sim 400$47

The central findings are that Hyak informed sampling captures more deaths than traditional cluster and stratified designs, and that Hyak combined with the spatial covariate model yields the smallest overall MSEs. The spatial model reduces bias by borrowing strength across neighboring villages, while variance is larger than in over-shrunk models; the net effect is lower MSE through a favorable bias-variance trade-off. The framework therefore defines an IMS not by cadence alone, but by persistent HDSS surveillance, adaptive sampling outside the surveillance core, and small-area estimation for mortality and cause-specific mortality fractions.

6. Acronym ambiguity, misconceptions, and recurring design principles

A common misconception is that “IMS” denotes a single standardized monitoring framework. The literature does not support that reading. In seismology, for example, IMS denotes the CTBTO International Monitoring System rather than Intensive Monitoring Survey. That network comprises more than 30 seismic arrays plus numerous 3-C stations and is used with waveform cross-correlation to construct automatic cross-correlation bulletins (XSEL). In this setting, the normalized correlation

$\sim 400$48

and a matched-filter statistic

$\sim 400$49

support low-magnitude event recovery below the International Data Centre detection level. The paper reports that WCC at arrays reduces detection thresholds by about an order of magnitude relative to beamforming and uses XSEL recurrence curves and Weak-to-Strict ratios to track precursory seismicity before the 2025 Kamchatka event (Kitov, 2 Jun 2026).

Once that acronym collision is separated out, the remaining IMS usages still do not define a single protocol. The 7DS IMS is an optical medium-band survey with daily cadence and five-year stacking; IMSNG is a telescope network optimized for hour-scale supernova catch; the ICU formulation centers on blurred video plus biosignals under hospital privacy rules; the multimodal indoor systems survey describes an IMS as a heterogeneous sensing-and-fusion architecture; and the Hyak formulation uses HDSS-centered informed sampling and spatial estimation for mortality surveillance (Ko et al., 26 Sep 2025, Im et al., 2019, Emberger et al., 2023, Nguyen et al., 2023, Clark et al., 2015).

This suggests a shared design pattern rather than a shared implementation. Across domains, the term is associated with repeated observation, intentional densification of information near decision-critical events, and explicit treatment of inferential failure modes. Those failure modes are domain-specific: the faint-$\sim 400$50 regime and low-$\sim 400$51/high-$\sim 400$52 degeneracies in 7DS; anisotropy, extinction, and host-selection bias in IMSNG; severe blur, occlusion, and a stationary-camera assumption in ICU monitoring; synchronization, demographic bias, and multi-user interference in indoor multimodal systems; and preferential sampling, HDSS-to-surroundings extrapolation, and verbal-autopsy misclassification in Hyak-based mortality monitoring (Ko et al., 26 Sep 2025, Im et al., 2019, Emberger et al., 2023, Nguyen et al., 2023, Clark et al., 2015).

Taken together, these literatures define IMS as a family of intensive-monitoring strategies rather than a single object. The exact meaning depends on the field, but the recurrent technical motif is the same: increase temporal density or event yield, preserve contextual information, and couple acquisition design to an estimator or inference engine that remains effective near the operational limit of the measurement system.