Sentinel 1 Time Series Analysis - Woodland Creation Grants
Tom Wilson
Overview
We analysed Sentinel-1 backscatter time series for conifer and broadleaved grant sites.
A single Sentinel 1 image offers little for detecting recent tree planting, but a 7-year time series shows interesting patterns.
We assessed whether two specific time series patterns are consistent across grant sites, enabling development of a monitoring process for Scottish Forestry.
By analysing a 7-year time series, we can consider if shorter time frames can meet business needs for this type of analysis in future.
Sentinel 1 - Synthetic Aperture Radar (SAR)
Forestry Grant Scheme Data
Downloaded from Scottish Forestry’s Open Data Portal.
Claim year 2018 coincides with first full year of JNCC’s Analysis Ready Sentinel 1 imagery.
Images were extracted and analysed for 1128 conifer grant schema polygons and 522 broadleaved.
Sentinel 1 ARD
Defra / JNCC Analysis Ready Sentinel 1 GRD imagery available on the CEDA Archive.
Details of the post-processing steps in user guidance.
First full year of Sentinel 1 ARD is 2018.
Images 1 Jan 2018 - 31 Dec 2024 used in this analysis.
How to Analyse Sentinel 1 GRD images
Use images captured in ascending and descending orbit separately.
Use both VV and VH polarisations.
Pixel values represent radar backscatter intensity in decibels (dB).
Surface roughness, vegetation structure, and moisture content influence backscatter.
Tree canopy produces higher backscatter in VV and VH than open land.
Sentinel 1 time series
- In total 12408 images were used from 2018 - 2024 from the Sentinel 1 JNCC ARD archive covering the forestry grant schemes selected.
| Year | S1 images used |
|---|---|
| 2018 | 2097 |
| 2019 | 2188 |
| 2020 | 2338 |
| 2021 | 2324 |
| 2022 | 1178 |
| 2023 | 1171 |
| 2024 | 1112 |
- From Dec 2021 Sentinel 1B satellite no longer functiong due to power supply issue. Sentinel 1C not launched until late 2024.
Conifer Time Series
The FGS claim polygons were used - explode multiparts and retain parts >= 0.5 ha.
Used 1128 conifer grant polygons of claim year 2018.
For each grant polygon, for each available Sentinel 1 image 2018 - 2024:
- Extract image pixel values intersecting the grant polygon to an array.
- Record the image date and orbit (ascending or descending).
- Take a median of pixel value array.
Output: A 2018 - 2024 time series of median pixel values for each grant polygon.
Monthly median - conifer
The noisy time series can be smoothed by taken a monthly median of the feature values.
Time series trends - conifer
Two trends were apparent in the time series for many 2018 claim sites:
A spike in VV, not VH seen before Spring 2020. Possibly ground prep for planting?
An overall upward trend in VV and VH backscatter, particularly in years 2020 - 2024. This is generally stronger in VH.
How to quantify the proportion of the 1128 (2018 claim) conifer grants showing these time series patterns?
Quantifying the time series trends - conifer
The aim is to quantify the two time series trends for all grants. Specifically, to determine:
How many of the 1128 conifer grant polygons with claim year 2018 show a spike in VV by March 2020, which may indicate ground preparation for tree planting.
A general increase in backscatter in both polarisations, but particularly in VH indicating trees are growing and a tree canopy is developing.
Identifying spike in VV early in time series (conifer ground prep)
- Using Z-scores to analyse the 2018 - 2024 time series of monthly medians.
\[ z = \frac{x - \mu}{\sigma} \]
where:
- \(x\) is the observed back scatter for a given month,
- \(\mu\) is the mean of the monthly backscatter values over the entire time series,
- \(\sigma\) is the standard deviation of these monthly backscatter values.
Z score threshold and assumptions - conifer
- A Z Score threshold of 2 was chosen.
- A VV time series spike, indicated by Z score of >= 2, should occur by spring 2020 (for our test).
- If a 2018 claim conifer site does not have this, indicates no ground prep for planting?
Z score threshold results - conifer
The z-scores were calcuated per grant polygon for VV, VH for ascending, descending orbits for the 2018 - 2024 time series of monthly medians.
The percentage of the 1128 conifer 2018 claim polygons having a z-score of 2.0 or greater (in either orbit) at three annual intervals to 31 March 2020 are shown below.
| Date to | VV Spike (%) | VH Spike (%) |
|---|---|---|
| 2018-03-31 | 11.44 | 3.72 |
| 2019-03-31 | 76.68 | 28.28 |
| 2020-03-31 | 83.95 | 34.22 |
Summary - Ground prep spike in VV - conifers
- 84% of 1124 conifer grants show the spike in VV within 2¼ years.
- Is this detecting ground preparation activities carried out before conifer planting?
- Consistent with VV polarisation being more sensitive to regular textures associated with man-made features, rather than vegetation.
- Need to repeat - compare non-planting, other years, understand timings of typical process on the ground.
Quantifying increase in backscatter over time series
- As the canopy develops, it introduces more volume scattering from branches and leaves - increased backscatter in the cross-polarised (VH) channel.
- Look at overall trend in backscatter time series separately from seasonal or short term variation.
- Generally higher backscatter in VV and VH in winter than summer, but not consistent.
- A 12-month rolling median was used to smooth the time series.
Applying linear regression - conifer
Linear regression was applied to assess the strength and statistical significance of (increasing) trend in backscatter.
Excluded first couple of years, so for 2018 claim, used 1 January 2020 to 31 December 2024.
Linear regressions were fitted to the smoothed VV and VH backscatter time series for each site (conifer grant polygon), and for each orbit (ascending and descending).
Linear regression results - conifer
- Percentage conifer 2018 claim sites showing a statistically significant (p < 0.05) positive trend in VV and VH backscatter since January 2020, along with the mean slope (dB/month) for those sites.
| Polarisation | % Sites +ve Trend (p < 0.05) | Mean Slope (+ve, p < 0.05) |
|---|---|---|
| VV | 50.98 | 0.017600 |
| VH | 90.16 | 0.024700 |
90% of 2018 conifer grants show significant positive trend in VH backscatter since January 2020 (51% for VV). VH has greater average slope than VV.
These results support the use of VH over VV for long-term monitoring of tree growth. Aligns with what is seen in other studies.
Broadleaves time series analysis
The same process applied to conifers was repeated for broadleaves / native broadleaves grants:
- 522 broadleaved grant polygons with claim year 2018 were analysed (after exploding multiparts and removing parts < 0.5 ha).
- 7 years of Sentinel 1 images 2018 - 2024 were extracted for these polygons as masked arrays.
- Time series were contstructed separately for ascending and descending orbits and then convered into monthly medians.
Observed trends - broadleaves
- The spike in VV polarisation - seen for conifers January 2018 to March 2020 - was not seen as frequently for 2018 claim broadleaved grants.
- A gradual increase in backscatter observed 2020 - 2024, particularly in VH polarisation. This was not as strong a trend as for conifers, probably because broadleaves take longer to become established and grow.
- Repeated conifer analysis for broadleaved 2018 grants.
- In future more customised methods could be investigated for broadleaves.
Z-score analysis - broadleaves
The conifer grant z-score analysis was repeated - looking for VV and VH time series spikes before 31 March 2020.
| Date to | VV Spike (%) | VH Spike (%) |
|---|---|---|
| 2018-03-31 | 12.84 | 2.49 |
| 2019-03-31 | 49.23 | 18.58 |
| 2020-03-31 | 63.98 | 27.97 |
VV polarisation the Z-score >= 2 spike was observed for 20% fewer sites for broadleaves (64% compared to 84% for conifers). Assume the same ground preparation not typical when planting broadleaves.
If so, a 20% difference in the occurence between conifer and broadleaved sites is perhaps not as great as expected.
Quantifying increase over the time series - broadleaves
Quantify the number of broadleaved grant sites showing an increase in VV and VH using linear regression.
The process described for conifers was applied to the broadleaved sites.
Firstly a monthly smoothing.
Linear regression - broadleaves
As for conifers, to quantify trend for all broadleaved sites, a linear regression was applied to the 1 January 2020 to 31 Dec 2024 portion of smoothed time series.
The table shows the proportion of the 522 broadleaved sites having a significant (p < 0.05) positive trend.
| Polarisation | % Sites +ve Trend (p < 0.05) | Mean Slope (dB / month, p < 0.05 and +ve) |
|---|---|---|
| VV | 56.51 | 0.0144 |
| VH | 82.18 | 0.0190 |
- Positive trend slightly less strong for broadleaves, probably because broadleaves grow more slowly.
Conclusions - Sentinel 1 Time Series Analysis
A time series analysis of Sentinel 1 GRD imagery applied to forestry grant schemes can be used to:
Show when and if ground prepartion has occurred by looking for a spike in the time series for VV polarisation within 1 - 2 years of the grant claim year.
Show if trees are growing on the site by applying linear regression to a smoothed time series over years 2 - 7 and looking for significant positive trends, particularly in VH polarisation. A minimum 4-5 years from claim year is needed.
Conclusions - Sentinel 1 Time Series Analysis
In future:
Business use case: Further understanding of forestry practices, details of business needs would help to refine analytical methods.
Ground truth data: Future analysis comparing a time series for non-planted sites with otherwise similar characteristics would be invaluable.
More advanced time series analysis and machine learning methods may be used once business requirements and ground truth patterns are understood.
Appendix - comparing VV spike with Planet
- VV spike occurring Nov 2018 - Jan 2019 for grant 18FGS27218.
- What can we see on optical 3 metre Planet images before and after this event?
Appendix - comparing VV spike with Google Basemap history
See freely available Google Basemap history for grant 18FGS27218.
