Intro: Crop segmentation

Sentinel-2 mission

Labels

Setting the task

• parcel-wise / pixel-wise?
• in-season?
• which year data can be accessed during training? (more questions)

Replicating the benchmark

• models
• training loops and details

Results + plots

• confusion matrix
• precision/recall
• patch-wise scores

Class distribution

(class distribution plot)

Class weights formula

• tuned the hyperparams to get $\alpha=0.1$

Grassland experiment

Data splits

Temporal generalisation experiment

• mention and quantify difference in train/test distributions

Map histogram

Temporal with 8 classes

In-season segmentation experiment

Data preprocessing

• patches are split into 6x6 subpatches

Experiment with subpatch center regions

Ignored class in the loss

Crop presence experiment

What is active sampling

• diagram: difference from active learning

Data flow diagram

Our strategies

• CCM
• Loss
• RHO-Loss
• random

Results

• training plot, confusion matrix

Learnings

Guidelines

Conclusion: summary of contributions