Step 3. Dieback detection
Step 3 : Detecting anomalies by comparing vegetation index to predicted value
The value of the vegetation index is compared to the vegetation index predicted from the periodic model calculated in the previous step, for each SENTINEL-2 acquisition available for the detection step. An anomaly is identified when the difference exceeds a threshold (in the expected direction in case of anomaly). For example, the CRSWIR is sensitive to canopy water content and tends to increase with decreasing water content.
Bark beetle outbreaks induce a decrease in canopy water content, therefore only CRSWIR values higher than expected values can be identified as anomalies. A pixel is detected as suffering from dieback when three successive anomalies are detected. This prevents false positive corresponding to one time events of anomalies due to an imperfect mask, or temporary climatic events.
As an example, the following figure shows the time series of the vegetation index along with the threshold for anomaly detection, and the date of detection :
Once anomalies confirmed (after three successive anomalies), pixel can return to normal state if no anomaly is detected for three successive acquisitions. This reduces the risk of false positive corresponding to long drought periods resulting in more than three successive anomalies. Also, information about those periods between false detection and return to normal, which we will call stress periods, can be stored.
A stress index can be computed for those stress periods, and for the final detection. It can be either the mean of the difference between the vegetation index and its prediction, or a weighted mean where the weight corresponds to the number of the date since the first anomaly of the period as illustrated in the following figure :
This stress index is meant to describe the intensity of detected anomalies in the period, and can be used as a confidence index for the final detection.
Comprehensive documentation can be found here.
Running this step using a script
Run the following instructions to perform this processing step:
from fordead.steps.step3_dieback_detection import dieback_detection dieback_detection(data_directory = data_directory, threshold_anomaly = 0.16, stress_index_mode = "weighted_mean")
Running this step from the command prompt
This processing step can also be performed from a terminal:
fordead dieback_detection -o <output directory> --threshold_anomaly 0.16 --stress_index_mode weighted_mean
NOTE : As always, if the model is already computed and no parameters were changed, the process is ignored. If parameters were changed, previous results from this step and subsequent steps are deleted and the model is computed anew.
The outputs of this step, in the data_directory folder, are :
- In the DataDieback folder, three rasters:
- count_dieback is the number of successive dates with anomalies
- first_date_unconfirmed_dieback : The date index of latest potential state change of the pixels, first anomaly if pixel is not detected as dieback, first non-anomaly if pixel is detected as dieback, not necessarily confirmed.
- first_date_dieback: The index of the first date with an anomaly in the last series of anomalies
- state_dieback is a binary raster with pixel as suffering from dieback (at least three successive anomalies) identified as 1.
- In the DataStress folder, four rasters:
- dates_stress : A raster with max_nb_stress_periods*2+1 bands, containing the date indices of the first anomaly, and of return to normal for each stress period.
- nb_periods_stress: A raster containing the total number of stress periods for each pixel
- cum_diff_stress: a raster with max_nb_stress_periods+1 bands containing containing for each stress period the sum of the difference between the vegetation index and its prediction, multiplied by the weight if stress_index_mode is "weighted_mean"
- nb_dates_stress : a raster with max_nb_stress_periods+1 bands containing the number of unmasked dates of each stress period.
- stress_index : a raster with max_nb_stress_periods+1 bands containing the stress index of each stress period, it is the mean or weighted mean of the difference between the vegetation index and its prediction depending on stress_index_mode, obtained from cum_diff_stress and nb_dates_stress The number of bands of these rasters is meant to account for each potential stress period, and another for a potential final dieback detection
- In the DataAnomalies folder, a raster for each date Anomalies_YYYY-MM-DD.tif whose value is 1 where anomalies are detected.