Lecture 2 - Multiple choice quiz | KIKompAG e-Learning website

(15 Questions - Approx. 15-20 minutes)

Instructions: Choose the best answer from the options provided.

Q1. What is the key benefit of self-supervised learning (SSL) in remote sensing for agriculture?

It allows full automation of data preprocessing
It requires no training at all
It learns from unlabeled imagery using pretext tasks
It reduces spectral resolution requirements

Q2. Which of the following SSL methods does NOT rely on contrastive learning?

SimCLR
MoCo
SimSiam
None of the above

Q3. Why are augmentations important in self-supervised learning?

They increase inference speed
They simulate label noise
They define the learning signal in pretext tasks
They reduce the dataset size

Q4. What makes VICReg different from other SSL methods?

It requires large labeled datasets
It avoids contrastive losses and focuses on redundancy reduction
It cannot handle multispectral data
It uses reinforcement signals

Q5. Which SSL model is designed to work well without negative samples?

MoCo
SimCLR
SimSiam
Barlow Twins

Q6. What is a practical benefit of integrating SSL with supervised fine-tuning in agriculture?

It speeds up satellite downloads
It ensures consistent cloud-free imagery
It improves model performance when labeled data is scarce
It eliminates the need for spectral calibration

Q7. Which of the following techniques can improve generalisation across regions and seasons?

SSL pretraining with fine-tuning
Simple supervised classification
Manual annotation only
Histogram equalization

Q8. In agricultural remote sensing, which type of augmentation is often avoided?

Color jittering
Horizontal flipping
Spectral mixing
NDVI scaling

Q9. How is SSL typically evaluated in low-label settings?

By visual inspection only
Using linear classifiers trained on top of frozen features
With random label shuffling
Based on augmentation performance

Q10. What does “representation collapse” refer to in SSL?

When models train too slowly
When all output features become similar or identical
When models use too many parameters
When datasets lack any variation