pat-s / 2019-feature-selection Goto Github PK

• Why use remote sensing data for forest health analysis?

• Show other studies using remote sensing on forest health

Initial issue: Some intermediate steps try to use some Sentinel scenes from a flight path which was not yet downloaded.

The reason for this is unclear.
Possible reasons:

• These images were added by ESA with a substantial delay.
• Recent changes in the PROJ library caused a slight projection shift of the area of interest which then intersected with new Sentinel scenes.

Downloading these new scenes errors at the moment due to an issue with the latest version of {getSpatialData}.
Updating the package was required because the version used until now did not work anymore due to changes in the Sentinel API.

Finish models for

• Laukiz 1 (1)

• Laukiz 2 (1)

• Luiando (1)

• Oiartzun (1)

• Supermodel (2)

(1) 5-fold 5 repeated CV
(2) Block CV (4 folds) on the plot level

Link https://pat-s.github.io/2019-feature-selection/report-defoliation.html is broken.

• Show all single plots and complete overview

• Describe the hyperspectral data

• Explain derivation of indices already here?

To see potential patterns

Characteristics

• Point density

• Coefficient of variation (COV) of defoliation

• more?

Allgemein

• Für alle plots (28)
• Für die gesamte study area
• 20 m resolution

Die prediction ist rein fürs Projekt, nicht fürs paper.
Prediction auf plot Ebene mit den hyperspectral Daten hab ich bereits gemacht: /mnt/mccoy/home/patrick/PhD/papers/03_hyperspectral/02_scripts/08_prediction.Rmd.

Wir nehmen xgboost mit 7 vars als Modell: /mnt/lossa/data_mccoy_kirk_scotty/patrick/mod/hyperspectral/prediction/xgboost_trained_tuned_7vars.rda

Sentinel-Daten: /mnt/lossa/data_mccoy_kirk_scotty/daphne/mod/Sentinel_clip

• Plot Namensgebung vereinfach/vereinheitlichen (in Skripten und Daten)

Sentinel Indices Ableitung

• Für alle verfügbaren Szenen eines Jahres von Apr - September ableiten
• Mittelwert nehmen
• Hoffentlich haben wir zumindest einen Wert pro Pixel
• Wenns nicht allzu kompliziert ist, wäre eine variable gut, die trackt, wie viele Szenen für ein Pixel letztendlich genommen wurden

After feature selection and hyperparameter tuning.

Maybe outside the CV with the models and indices from the CV. Advantage: We can parallelize it then and have a break (in case sth goes wrong in the CV).