Quick wins for a faster PC:
Repair Windows errors before they cause bigger problemsFix Now →Fix the driver behind crashes, sound loss and screen glitchesFind Drivers →iTechGuides is reader-supported. When you buy through links on our site, we may earn an affiliate commission. As an Amazon Associate I earn from qualifying purchases. Learn more
Desmond Kangah’s work uses Sentinel-1 radar satellite data to measure how the ground in East Baton Rouge Parish, Louisiana, is moving, then applies machine learning and a forecasting model to show where that movement is concentrated and where it may continue. Two 2026 studies describe the approach: his Spring 2026 civil engineering thesis at Louisiana State University, and a co-authored paper in the IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing (JSTARS) focused on transportation infrastructure. Both identify patterns that point to where people should look more closely. Neither establishes that a particular road or bridge is safe or unsafe.
The four-step workflow
Kangah’s approach moves through four stages. Each stage answers a different question, and the outputs of one stage are not the same kind of evidence as the next.
- Measurement. Sentinel-1 radar images of the same ground are collected repeatedly over time. Interferometric processing compares those images to estimate how the land surface has moved between acquisitions, producing a time series for each measurement point. The thesis reports 246 ascending-track acquisitions. The EGU conference abstract for the same work describes Sentinel-1 Small Baseline Subset (SBAS) time series covering 2017 to 2025.
- Interpretation. Ensemble machine learning models are trained to map where subsidence is concentrated and how susceptible an area is. Explainable AI methods then show which conditions drive those results, such as fault proximity or groundwater withdrawal.
- Forecast. A physics-constrained long short-term memory (LSTM) network projects future deformation rates. In the thesis, these are model projections through 2030, not observations of what has already happened.
- Targeted follow-up. Flagged areas are the places where engineers would direct independent checks and site inspection. The studies treat this step as the job of people, not of the map.
InSAR, short for interferometric synthetic aperture radar, measures surface movement over repeated observations. It does not see cracks, measure load on a structure, or tell an engineer what is happening inside a pier or embankment. That distinction matters for everything that follows.
Two studies, two methods
The thesis and the transportation paper share a study area, a satellite sensor, and a general goal. They differ in the InSAR method, the models, and the question each one answers. They should not be read as the same analysis reported twice.
#1 Best Overall
| Element | LSU thesis (Spring 2026) | IEEE JSTARS paper (2026) |
|---|---|---|
| Study area | East Baton Rouge Parish, Louisiana | East Baton Rouge Parish, Louisiana |
| Satellite data | Sentinel-1 | Sentinel-1 |
| InSAR method | Small Baseline Subset (SBAS) | Persistent-scatterer InSAR (PSI) |
| Models and explanation | Ensemble machine learning, physics-constrained LSTM, explainable AI | Random Forest (RF) and SHapley Additive exPlanations (SHAP) |
| Primary question | Where subsidence is concentrated and how it may evolve through 2030 | Localized deformation and stability near major interchanges and bridges |
| Acquisition count | 246 ascending-track acquisitions | Not stated in the paper’s available abstract text |
| Main reported drivers | Spatial concentration near fault structures and areas of intensive groundwater withdrawal | Fault proximity and lithologic variability as dominant controls; precipitation as a secondary contributor |
The distinction between methods matters for interpretation. SBAS, used in the thesis, is built for mapping deformation patterns over broad areas and feeding them into a temporal model. PSI, used in the paper, tracks stable reflective points on the ground, which suits assessing movement at specific built features such as bridges and interchanges. The two outputs answer different questions and should not be presented as interchangeable.
What the thesis reports
Where the ground is moving
The thesis reports subsidence that is spatially concentrated near fault structures and in areas of intensive groundwater withdrawal within East Baton Rouge Parish. The concentration is the finding; the thesis does not present it as a universal cause-and-effect relationship across all locations.
How well the models fit the data
The thesis compares several models. The figures below are the thesis’s own reported values. They describe model performance in the thesis’s evaluation, not accuracy at any particular bridge or intersection.
| Model | Reported R² | Reported AUC | Other reported value |
|---|---|---|---|
| Extra Trees | 0.92 | 0.97 | Not stated |
| Random Forest | 0.88 | 0.94 | Not stated |
| Physics-constrained LSTM | 0.83 | Not applicable | Pearson r = 0.92 |
The ensemble models are described as susceptibility classifiers, which is why AUC is reported for them. The LSTM is a forecasting model and is evaluated with regression measures.
The 2030 projection
The thesis’s forecast projects a mean velocity of −0.53 mm/year, accumulating to −6.42 mm over the five-year forecast period through 2030. These are the model’s projections. They are not measurements of deformation that has occurred after the study. The cumulative figure is not a simple five-year multiple of the mean rate; five years at −0.53 mm/year would be about −2.65 mm. The sources available for this article do not describe how the cumulative value was constructed, so it should be read as the model’s own output rather than a derived product of the mean velocity.
What the transportation paper reports
Localized signals near interchanges and bridges
The paper reports localized deformation near major interchanges and bridges in the parish. Signals are stronger in fault-bounded corridors and in areas where the subsurface is stratigraphically variable, meaning the soil and rock layers change over short distances. The paper’s own framing is that these are areas that warrant attention, not confirmed hazards.
Which factors the model ranks highest
Using Random Forest with SHAP explanations, the paper identifies fault proximity and lithologic variability as the dominant controls on deformation susceptibility. Precipitation is identified as a secondary contributor. SHAP shows how much each input pushes a prediction up or down, which is useful for checking whether the model is responding to physically plausible conditions. It does not prove that any one factor causes movement.
The Tool Desk
Outbyte Driver Updater FREEFix the driver behind crashes, sound loss and screen glitchesFind Drivers →Outbyte PC Repair FREEClear out junk files and repair common Windows errorsFree Scan →The GNSS check
The paper compares its InSAR time series with independent GNSS observations from reference stations. This is a consistency check: it shows whether the satellite-derived movement agrees with a separate measurement method at the locations tested. It does not validate every pixel or every infrastructure asset in the study area.
The paper’s own summary of its contribution reads: “By combining PSI, RF, and SHAP, the framework enhances interpretability, improves deformation-susceptibility mapping, and supports infrastructure risk assessment.” (Ahmed Abdalla, Desmond Kangah, and Abdelrahim Salih, IEEE JSTARS, 2026.)
Rank #4
Reading the numbers without overreading them
- Projections are not observations. The 2030 values are what the model expects under its assumptions. Measured deformation from the satellite time series is a different kind of number.
- Susceptibility is not cause. A map that shows an area resembles places where deformation has been detected tells you where to look, not why the ground is moving there.
- Fit scores have a scope. An R² of 0.92 describes how well a model reproduced the thesis’s data. It does not guarantee the same accuracy in a different parish, with different data, or at a single structure.
- Agreement with GNSS is partial. Consistency at the points compared supports the method. It is not a blanket statement about every location.
- Method outputs differ. A PSI point’s movement, an SBAS area’s trend, and a model’s susceptibility score are three different products, even when they cover the same street.
What these studies do not establish
The thesis and the paper describe analysis methods and their results for East Baton Rouge Parish. They do not describe an operational warning service, a deployed monitoring system used by a transportation agency, or any measured outcome in which the approach prevented damage. Whether the workflow would perform the same way in other parishes or on other datasets is not tested in these sources.
How the outputs fit an inspection workflow
The studies do not prescribe an inspection procedure. The sequence below is one way a transportation engineer could use these kinds of outputs, drawn from the logic of the studies themselves.
Quick Recap
- Review the InSAR time series for the corridor or structure of interest, noting where movement is persistent rather than a single anomaly.
- Check the flagged locations against an independent measurement, such as GNSS or conventional survey data, before drawing conclusions.
- Compare the pattern with mapped fault locations, subsurface variability, and recent precipitation records to see whether the pattern is consistent with the drivers the paper identifies.
- Prioritize the locations for site inspection, where condition is observed directly.
- Let the engineering judgment of qualified staff determine any action. The model output is one input to that judgment.
Where to read the sources
- LSU thesis record, Spring 2026: https://repository.lsu.edu/gradschool_theses/6366/. The repository record is the abstract. The thesis file is listed for download on March 26, 2029, so the full text is not currently downloadable from this record.
- IEEE JSTARS paper, 2026: Abdalla, Kangah, and Salih, “PS-InSAR and Explainable Machine Learning for Local Transportation Infrastructure Stability in Baton Rouge.” DOI 10.1109/JSTARS.2026.3705283. Published June 22, 2026; current version July 28, 2026. The full text is available through ResearchGate at https://www.researchgate.net/publication/407497573_PS-InSAR_and_Explainable_Machine_Learning_for_local_Transportation_Infrastructure_Stability_in_Baton_Rouge.
- EGU26 conference abstract: Kangah and Abdalla, EGU26-16451, updated March 14, 2026. This abstract describes the 2017–2025 SBAS time series and the forecast’s quantified uncertainty, but it does not report the thesis’s specific forecast values. https://meetingorganizer.copernicus.org/EGU26/EGU26-16451.html
- Louisiana Transportation Conference 2025 program: documents an earlier poster on transportation deformation monitoring using InSAR. https://www.ltrc.lsu.edu/ltc_25/pdf/LTC_2025_program_online.pdf
- Background profile (September 24, 2026): https://techbullion.com/how-desmond-kangah-uses-geoai-to-monitor-land-subsidence-and-protect-transportation-infrastructure/. This profile is useful background, but the technical claims in this article rest on the thesis, the paper, and the EGU abstract.
“
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
