This blog post qualitatively compares the OPERA Sentinel-1 disturbance product with the OPERA DIST-HLS products across a few specific sites. In performing this comparison, we provide a perspective on the capabilities, strengths, and limits of the Sentinel-1 product. The DIST-HLS has two disturbance products: the generic product (GEN-STATUS) and vegetation product (VEG-STATUS). The DIST-S1 is a generic change product and we refer to it as DIST-S1 hereafter. The DIST-HLS VEG-STATUS first models “vegetation cover” and then delineates changes when a pixel undergoes vegetation cover loss. We highlight that vegetation gain is not considered. The DIST-HLS GEN-STATUS is the statistical deviation of the recent acquisition with respect to the baseline of the images capturing generic change. While both DIST-HLS disturbance products are validated carefully, there are no formal requirements for the GEN-STATUS. The DIST-S1 product delineates disturbances based on statistical variations in a very similar fashion to DIST-HLS’s GEN-STATUS. When comparing the above products (DIST-S1, VEG-STATUS, GEN-STATUS), we share all three product layers.

Important Thematic Differences between DIST-S1 and DIST-HLS

Before jumping in to the interactive maps, we want to highlight a few items that will be prevalent across all comparisons.

• Statistical measures of disturbance (i.e. DIST-S1 and GEN-STATUS) will see any deviation from the mean as a disturbance. Vegetation loss, however, measures a specific type of disturbance. For example, generic changes should capture vegetation growth or structural changes within a cleared area. However, vegetation loss delineations (e.g. VEG-STATUS) will not delineate these disturbances.
• The High/Low alerts for statistical measures of disturbance (i.e. DIST-S1 and GEN-STATUS) indicate the likelihood of disturbance. That is to say, a “high confidence” alert is much less likely than a “low confidence” alert, where we are analyzing these probabilities using the baseline images. Reframing using the complement, a “high confidence” alert is more likely to be an anomaly than a “low confidence” alert. For the VEG-STATUS, there is a thematic view of “high” and “low”: “High” indicates over >50% vegetation cover loss in a pixel and low indicates <50%. In particular, statistical “high”/“low” have different meanings than those in VEG-STATUS. For example, a disturbance that removes <50% vegetation cover could be viewed as a highly unlikely with respect to a statistical measure.
• The DIST-S1 product uses a transformer model to estimate the mean and standard deviation of the set of baseline imagery. At a high level, the DIST-S1 model is accounting for the temporal ordering of the baseline imagery and therefore more recent images are more likely to contribute to accurate parameter estimates than those images acquired further in the past. The VEG-STATUS estimates the baseline vegetation cover by compositing all of the baseline images together and estimating the vegetation cover from the composite. Similarly, the GEN-STATUS estimates per-pixel normal parameters via standard sample estimation. In particular, the DIST-HLS products do not consider temporal ordering.
• While the disturbance delineations may align in the different scenarios, the acquisition mode of sidelooking SAR fundamentally contrasts the down-looking optical data and a disturbance area in SAR can appear to be “smoothed” compared to HLS.
• DIST-HLS has a sampling rate of 2-3 days whereas DIST-S1 has a sampling rate of about 1 week. However, Sentinel-1 does not regularly distribute all the data that it can, except over Europe and parts of North America. That said, there are instances of high latitude where scenes are too dark for HLS to distribute. Additionally, when comparing and confirming disturbances delineations from Sentinel-1 over time via different acquisition geometries (i.e. ascending and descending passes), it is harder to confirm (i.e. verify over time) disturbances particularly if disturbances are visible in one acquisition geometry and not the other.
• If there is alignment for a particular disturbance, HLS is often able to detect “finer” features than SAR, i.e. features that are 1-2 pixels in width like a clearing for a road or path in a forest. One reason is the geometry of acquisition elaborated on in the previous bullet (vertical vs. side-looking).
• SAR is sensitive to structure as opposed to spectral changes (like “green”-ness). For example, sparse green vegetation that is cleared will not appear as disturbance in SAR and steep inclines will sometimes prevent accurate delineations from being identified on such terrain.
• SAR is sensitive to soil moisture, rain, snow, and floating vegetation/sediment. If these appear in a recent acquisition and represent deviations from the curated baseline, these disturbances will show up as provisional changes. Through the process of confirming disturbances (i.e. accumulating disturbances through time), these disturbances will go away.
• The confirmation process provides a means to track disturbances that are consistent in time and accumulate these disturbances in the recent product. However, this process also allows for “reset” of disturbances: if a change is “confirmed” and then returns to it’s baseline as “finished”, it can “reset” if a new change is detected after the finished status. If this cycle occurs, then changes that were confirmed in time are not accumulated. For DIST-S1, where regrowth, urban development, and moisture sensitivities are delineated, there can be resets that makes the products appear different particularly VEG-STATUS and DIST-S1.

Dr. Amy Pickens selected many of the sites below for informal comparisons except those related to a specific hazard. Most of the data represents changes detected in products generated on or before August 1st, 2025. We used 4 months of data products prior to a given date to confirm the changes in the products shown.

Examples

Logging

Central Sweden

Below is a portion of a managed forest in Gävleborg County, Sweden.

Things to note:

• There is a lot of alignment between VEG-STATUS and DIST-S1. We note that the large “high” confidence alerts in DIST-S1 are “low” alerts in VEG-STATUS. This means that this was “unlikely” with respect to the baseline collected from S1, but the modeled vegetation cover within said pixels is <50%.
• The larger disturbances delineated in DIST-S1 not present in VEG-STATUS appear over bare areas with respect to the base satellite imagery. We speculate such areas corresponds to recent vegetation growth that is by definition not in VEG-STATUS.
• The DIST-HLS products contain some finer features like trimming forest stands at the edges.

Mining

Indonesia

Below is a nickel mine on Weda Bay in Indonesia.

Things to note:

• The process of mining in this area undergoes the cycle: vegetation clearing and then structural build-up. We note that statistical/generic measures will detect both, but vegetation loss delineations (VEG-STATUS) will only detect the first.
• We can see build up of various mining areas switching between the basemaps (ESRI and Google Satellite).

Development and Road Expansion

Northeastern China

This scene is capturing some of the recent development that is going on throughout China on a regular basis. This scene shows the conversion of green areas to buildings with solar paneled roofs as well as road expansion.

Things to note:

• We see some of the development at high resolution looking at the Google Maps and then switching to ESRI Imagery.
• The DIST-HLS product shows the creation of a narrow road or railway being built. However, these are recent/ongoing developments and the changes are not confirmed (i.e. provisional) in either VEG-STATUS or GEN-STATUS. In the high-resolution imagery, you can see the highway piles being laid without highway atop them yet.
• The DIST-S1 product is very good at seeing the construction of buildings with solar paneled roofs.

Balitomore, Maryland

This shows the port of Baltimore and the surrounding industrial area.

Things to note:

• Large objects that move into the scene will be delineated as disturbance. Physically, this is the result of the change from specular reflection (flat ground or water) to double bounce (large objects):
  • Cargo ships that park in the various loading docks (assuming they were not present in the baseline) appear as disturbance in the DIST-S1 product. You can see some of these massive ships by switching the satellite basemaps (from Google to ESRI).
  • Large cargo areas often get shown as disturbances due to the moving and clearing of large storage containers in the scene.
  • A few massive parking lots appear as disturbances either because they are new (after construction atop clear land) or the more regular occurences of parked cars. Both would change the returns from specular to double bounce.
• The DIST-HLS and DIST-S1 agree on changes that concern the clearing of vegetation (e.g. trees) though DIST-S1 clearly show far more disturbances in this coastal urban scene. It’s also worth noting that the F1 mask excludes water from being considered in disturbances (and the baseline).

Shifting Cultivation

Laos

Below we see the shifting cultivation of Rice Paddies in Southern Laos.

Things to note:

• There is some hilly rice paddies that can be seen being created by switching between ESRI and Google Satellite Imagery in the large areas of disturbance towards the center of the imagery.
• There appears to be good alignment between VEG-STATUS and DIST-S1 - however, DIST-S1 is indicating “finished” disturbances (white) whereas VEG-STATUS has ongoing confirmed disturbances (red). This may be the result that clearing the vegation for growing the new rice paddies is a change but once the rice paddies are mature with vegetation, they appear similar to the baseline SAR images. However, from the point of view of VEG-STATUS the clearing is a permanent decrease in vegetation cover because rice paddies are not as vegetated as they were with respect to the baseline.