|
| 1 | +# Background |
| 2 | + |
| 3 | +This page shows scientific background information about the workflow. |
| 4 | +Further information is provided in the publication [(Cremer et al. 2020)](https://doi.org/10.1109/JSTARS.2020.3019333). |
| 5 | + |
| 6 | +## Introduction |
| 7 | + |
| 8 | +Tropical forests help stabilize the global climate and protect biodiversity. |
| 9 | +Remote sensing technologies, like optical sensors, have been used to monitor these forests, but they struggle with cloud cover, especially during rainy seasons. |
| 10 | +Synthetic Aperture Radar (SAR) can penetrate clouds and is used to map forest changes. |
| 11 | + |
| 12 | +SAR data can vary due to environmental conditions, thus using time series data helps improve accuracy. |
| 13 | +Current methods for detecting deforestation with Sentinel-1 data fall into two groups: one uses single or few images to detect changes, while the other uses time series to predict changes. |
| 14 | +Various techniques, like using radar shadows or Bayesian updating, help identify deforestation. |
| 15 | + |
| 16 | +## Methods |
| 17 | + |
| 18 | +The underlying algorithm is based on Recurrence quantification analysis (RQA). |
| 19 | +Hereby, one counts signal change across every possible time point pair. |
| 20 | +Recurrence plots visualize all comparisons: |
| 21 | + |
| 22 | + |
| 23 | +Recurrence plots for (a) the sum of two sine waves with different frequencies, (b) a step function with noise, and (c) a sine wave with trend. (b) shows a deforestation event whereas the other two just show seasonal effects or trends. |
| 24 | + |
| 25 | +A location is considered as deforested if a change was detected often enough (threshold approach). |
| 26 | +In particular, the TREND metric is used to make this decision: |
| 27 | + |
| 28 | +$$ |
| 29 | +\text{TREND}= \frac{\sum _{\tau =1}^{\tilde{N}}(\tau - \tilde{N}/2)(\text{RR}_\tau - \langle \text{RR}_\tau \rangle)} |
| 30 | +{\sum _{\tau =1}^{\tilde{N}}(\tau - \tilde{N}/2)} |
| 31 | +$$ |
| 32 | + |
| 33 | +Hereby, $\tilde{N}$ is the number of time steps and $RR_i$ is the number of recurrent values on the ith diagonal. |
| 34 | +It represents the linear regression coefficient over the $RR$ of the diagonals in comparison to their distance to the main diagonal. |
| 35 | +It indicates whether the process is drifting or not. |
| 36 | +The $TREND$ value is computed for each location individually and reported in the final result data cube. |
| 37 | + |
| 38 | + |
| 39 | +## Data |
| 40 | + |
| 41 | +Up to now, this algorithm was only applied to Sentinel-1 Sigma Nought backscatter data as processed using [Wagner et al. 2021](https://www.mdpi.com/2072-4292/13/22/4622) in [Equi7Grid projection](https://github.com/TUW-GEO/Equi7Grid). |
| 42 | +Using less corrected imagery may be possible but has not been tested yet. |
| 43 | + |
| 44 | +## Parameters |
| 45 | + |
| 46 | +First, the workflow is executed on a given spatiotemporal extent. |
| 47 | +The spatial extent doesn't influence the result of individual locations, since neighboring pixels are not considered. |
| 48 | +For the Sentinel-1 Sigma Nought dataset in Equi7Grid projection, the area to be analyzed is given as a list of tiles. |
| 49 | +The time span should be a multiple of a year to reduce potential seasonal bias. |
| 50 | +The algorithm itself can be tuned by setting the threshold value that determines whether to count a time point par change or not. |
0 commit comments