Container dwell-time distribution at ports as trade-finance risk indicator
Repeat very-high-resolution optical imagery of container terminals lets analysts track individual stack persistence across successive passes, turning empirical dwell-time distributions into early indicators of port congestion, customs delays and sanctions-related cargo holds that matter directly to trade-finance risk.
Sensors
- Planet SkySat: 50 cm native resolution (after pan-sharpening), capable of multiple collects per day over a tasked location. At 50 cm, a standard 20-foot container (6.1 m × 2.4 m) subtends roughly 12 × 5 pixels, sufficient to distinguish individual boxes in a stack row. Revisit at sub-daily cadence when multiple satellites are tasked in train.
- Maxar WorldView Legion: Designed for 30 cm resolution and up to 15 revisits per day over a target site when the full constellation is operational. That cadence is the closest commercial optical imagery comes to a persistent stare, which matters for short-dwell containers that may move within hours.
- Airbus Pléiades Neo: 30 cm resolution, stereo and tri-stereo collection in a single pass. The stereo capability allows photogrammetric estimation of stack height, which is a proxy for the number of container tiers and therefore the volume of cargo sitting idle. Revisit is roughly once daily per satellite, with two satellites in the constellation.
- Satellogic NewSat: 70–100 cm resolution at nadir, with hyperspectral capability on some satellites. Resolution is marginal for individual container discrimination at crowded terminals but adequate for block-level change detection and stack-area estimation. The constellation's commercial model emphasises archive depth and tasking flexibility.
What a stationary container actually tells a lender
A letter of credit is a time-bounded instrument. The issuing bank's exposure is live from the moment the credit opens until the cargo is released and title transfers. If containers linked to a financed shipment sit on the apron for days beyond the contractual free-time window, the bank faces demurrage accrual, potential cargo deterioration and, in sanctions contexts, the possibility that the goods are frozen rather than delayed. Dwell time is not an abstract operational metric; it is a direct expression of credit risk.
The key insight is that a container that does not move between two satellite passes is generating a detectable signal. Aggregate enough of those non-movement observations across a terminal and you get an empirical dwell-time distribution: a histogram showing what fraction of containers have been stationary for one day, three days, a week, and so on. Shift in the tail of that distribution, particularly an elongation beyond the typical free-time threshold of three to five days, is the early warning that something has changed at that port.
From pixels to persistence: the detection chain
The photogrammetric pipeline begins with orthorectification. Container terminals are flat, which simplifies the geometry, but crane booms and stack shadows introduce occlusion that must be masked before any change-detection step. At 30–50 cm resolution, a standard TEU (twenty-foot equivalent unit) is large enough to be resolved individually in a clear, nadir-looking collect. Stacks of four or five tiers appear as a single rectangular object; the shadow length and the stereo parallax from a Pléiades Neo tri-stereo pass can estimate stack height to within roughly one container tier under good conditions.
Change detection between successive orthorectified images uses either image differencing on aligned chips or feature-matching on detected container objects. The latter approach, sometimes called object-based image analysis (OBIA), segments the terminal into candidate container objects, assigns each a colour signature and positional centroid, then attempts to match objects across time steps. A container that matches in position across N consecutive passes is assigned a minimum dwell of N × revisit interval. The word 'minimum' matters: if the container was already present before the archive window opens, the true dwell is longer than observed.
Colour is a useful but imperfect discriminator. Containers of the same operator share livery, so colour alone cannot uniquely identify a box. Positional persistence within a stack row is the primary signal; colour consistency is a supporting check. At 50 cm resolution, painted identification numbers on container tops are not reliably legible, so the method operates at the statistical level of stack blocks rather than individual unit tracking.
Cloud, shadow and the gaps you must model
Optical imagery fails in cloud. Major container hubs in South and South-East Asia, West Africa and the Gulf of Mexico sit under cloud cover for significant fractions of the year. A terminal in Singapore or Colombo may be obscured for three to five consecutive days during the south-west monsoon. Any dwell-time model that treats a cloudy gap as 'no change' will systematically overestimate dwell; any model that discards the observation entirely loses statistical power precisely when port activity may be elevated.
The honest treatment is to model cloud gaps explicitly as censored observations in a survival-analysis framework. A container observed at time T1, obscured for two days, then observed again at T4 has a censored interval. Maximum-likelihood estimation under a parametric dwell-time distribution (log-normal or Weibull are common in port-operations literature) can recover the underlying distribution even with substantial censoring, provided the censoring mechanism is independent of the dwell process itself. That independence assumption breaks down if, for example, a port authority systematically moves containers during the hours that happen to be cloud-free, which is unlikely but worth noting.
Crane shadows are a separate, more tractable problem. A gantry crane at a major terminal casts a shadow tens of metres long at low sun angles, occluding the containers beneath. Shadow masks derived from the solar geometry at acquisition time allow affected pixels to be flagged rather than misclassified as empty apron.
Translating dwell distributions into credit-risk signals
A single port observation is almost meaningless without a baseline. The useful product is the deviation of the current dwell distribution from the port's own historical norm, calculated separately for each terminal zone (import apron, export apron, transshipment yard) because their dwell dynamics differ structurally. An import apron where the 90th-percentile dwell has risen from four days to nine days over three weeks is a concrete, quantified signal that customs processing has slowed or that a specific commodity category is being held.
For trade-finance applications, the signal can be linked to specific vessel calls. If AIS data places a vessel at berth during a given window and the import apron adjacent to that berth shows an anomalous dwell spike in the weeks following, the analyst has a spatially and temporally coherent chain of evidence. This does not prove that a specific financed cargo is affected, but it raises the probability sufficiently to justify a compliance check or a covenant conversation with the borrower.
Sanctions screening adds a further dimension. Ports known to handle transshipment for sanctioned entities can be monitored for unusual dwell patterns that may indicate cargo holds pending regulatory review. The imagery does not identify the beneficial owner of a container, but it can flag the terminal-level anomaly that prompts a deeper document review.
Honest limits of the method
Resolution sets a hard floor. At 50 cm, individual containers are resolved but not uniquely identified by serial number. At 1 m resolution, stack rows are detectable but individual box boundaries blur. Anything coarser than 1.5 m is unlikely to support container-level analysis at a densely packed terminal; it reverts to block-level area estimation, which is the domain of the sibling page on port throughput estimation.
Revisit cadence determines the minimum detectable dwell. A container that arrives and departs within a six-hour window between two SkySat passes leaves no trace in the imagery. The method is therefore blind to rapid transshipment and is most sensitive to dwell times exceeding the revisit interval by a comfortable margin, typically 24 hours or more.
Finally, the method produces distributional signals, not individual cargo tracking. A lender seeking to verify the status of a specific container identified by its booking number cannot do so from optical imagery alone. The product is a port-level or terminal-zone-level risk indicator, best used as a screening tool that directs human investigation rather than as standalone evidence.
Putting it into practice
A workable monitoring programme combines commercial tasking on a priority terminal list with open-archive imagery from Sentinel-2 (10 m resolution, 5-day revisit) as a coarse continuity layer during cloud gaps. The Sentinel layer cannot resolve individual containers but can detect gross changes in yard utilisation through spectral reflectance shifts on the apron surface. The commercial layer provides the high-resolution dwell statistics; the open layer fills temporal gaps in the censoring model.
Satellize structures analytics engagements of this type as periodic scored reports delivered alongside a GIS layer showing terminal-zone dwell anomaly scores, updated on a cadence matched to the client's credit-review cycle. The approach is similar in spirit to the crop-estimation methodology Satellize runs for the Kingdom of Tonga, adapted from agricultural field persistence to industrial yard persistence. The underlying statistical discipline is the same: repeated observation, honest accounting of gaps, and a distributional output rather than a false-precision point estimate.
Banks and trade-credit insurers considering this capability should begin with a pilot on two or three terminals where they already hold exposure and where ground-truth from port agents is available to validate the dwell estimates. That calibration step is not optional; without it, the anomaly thresholds are guesses.
Typical figures
| Spatial resolution (commercial) | 30–50 cm (Pléiades Neo, WorldView Legion, SkySat); 70–100 cm (Satellogic NewSat) |
| Spatial resolution (open archive fill) | 10 m (Sentinel-2 MSI); sufficient for yard-area utilisation only, not individual containers |
| Revisit cadence | Sub-daily to once daily for commercial tasking; 5 days (Sentinel-2, mid-latitude, single satellite); WorldView Legion designed for up to 15 passes per day at full constellation |
| Minimum detectable dwell | Approximately equal to the revisit interval; containers cycling faster than 6–12 hours between passes are not reliably detected |
| Minimum resolvable target | Individual 20-ft container (6.1 m × 2.4 m) at 50 cm resolution; stack rows at 1 m resolution |
| Spectral bands used | Panchromatic and natural-colour RGB for object detection; NIR band assists shadow and crane-boom masking |
| Stack height estimation | Photogrammetric stereo (Pléiades Neo tri-stereo): ±1 container tier under good conditions |
| Cloud impact | Systematic gaps; tropical and monsoon-affected ports may lose 3–5 consecutive days. Modelled as censored observations in survival-analysis framework |
| Archive depth | Commercial constellations: 3–6 years depending on operator and terminal priority. Sentinel-2: from 2015 (open access) |
| Delivery formats | GeoTIFF orthorectified imagery, GeoJSON terminal-zone polygons, CSV dwell-score time series, PDF scored report |
Analytics Satellize can run
| Terminal-zone dwell anomaly score | Object-based image analysis (OBIA) change detection across co-registered successive collects; deviation from rolling historical baseline | GIS layer (GeoJSON) with per-zone anomaly scores, updated on agreed cadence |
| Empirical dwell-time distribution | Survival analysis (Weibull or log-normal parametric fit) on stack-persistence observations with censored intervals for cloud gaps | CSV time series of distribution parameters (median, 90th percentile, tail weight) per terminal zone |
| Stack-height and yard-utilisation estimate | Photogrammetric stereo processing (Pléiades Neo tri-stereo) for tier count; apron-coverage ratio from segmented container-object area vs total apron area | Tabular report with stack-height map overlay (GeoTIFF) |
| Vessel-call-linked dwell spike report | Spatial and temporal join of AIS vessel-call records to post-berth apron dwell anomaly; flagging of statistically significant co-occurrences | PDF scored report per vessel call of interest, with annotated imagery |
| Sanctions-port monitoring alert | Threshold-based alerting on dwell-distribution tail elongation at designated high-risk terminals; change-point detection on rolling 30-day baseline | Email or API alert with supporting annotated imagery and zone-level dwell statistics |
| Cloud-gap-corrected dwell confidence interval | Maximum-likelihood estimation under censored survival model; Monte Carlo uncertainty propagation across cloud-gap durations | Confidence-interval bounds appended to dwell-score CSV; methodology note in PDF report |
Who does the work
We can get this done for you. Satellize runs its own analyst desk and a strong science team. You do not buy a data feed and work out what it means; our people source the imagery, run the analysis described on this page, and hand you the answer with its confidence limits stated. Discuss this requirement.