Lecture 1 - 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 main limitation of traditional deep learning models in agricultural Remote Sensing tasks?

They require GPU processing
They depend on large volumes of labeled data
They cannot process multispectral imagery
They are limited to urban areas

Q2. Which of the following is NOT a small-data strategy described in the lecture?

Few-shot learning
Active learning
Reinforcement learning
Self-supervised learning

Q3. Why is spatial k-fold cross-validation recommended in geospatial applications?

It increases training speed
It allows for random sampling of validation data
It ensures that training and test data are spatially independent
It balances class distribution across folds

Q4. Which technique involves training a model using noisy or incomplete labels?

Weakly supervised learning
Zero-shot learning
Transfer learning
Process-aware learning

Q5. In which scenario is zero-shot learning most useful?

When pretraining is not feasible
When new, unseen classes need to be predicted
When all labels are available but noisy
When training with satellite video data

Q6. Which of the following techniques is most suitable when no labeled data is available at all?

Semi-supervised learning
Self-supervised learning
Active learning
Weakly supervised learning

Q7. Spatial k-fold cross-validation is superior to random splits in geospatial tasks because:

It uses less training data
It increases model accuracy
It separates training and validation areas geographically
It allows multiple model types

Q8. Few-shot learning is used when:

You have multiple labels per image
Training data is noisy
You need to learn from very few samples per class
A model is pretrained on the same data

Q9. Process-aware learning integrates:

Metadata from satellite sensors
External ecological or physical models
Automatic feature generation
Transfer weights from natural image models

Q10. What is the primary purpose of active learning in small-data scenarios?

To synthesize new labeled samples
To transfer knowledge between models
To identify and annotate the most informative unlabeled samples
To apply physics-based constraints to the learning process

Q11. Which technique can reduce the need for a large number of labels by learning from the relationships between different tasks?

Semi-supervised learning
Multitask learning
Ensemble learning
Few-shot learning

Q12. What distinguishes semi-supervised learning from supervised learning?

It uses synthetic data instead of real imagery
It trains models only on test data
It combines labeled and unlabeled data during training
It requires no human input

Q13. Which strategy is most commonly used to initialize a model with pretrained weights from a different domain?

Ensemble learning
Transfer learning
Active learning
Zero-shot learning

Q14. What is one benefit of using ensemble learning in agricultural remote sensing tasks?

It guarantees perfect accuracy
It removes the need for data preprocessing
It provides more stable predictions by combining multiple models
It can only be used with hyperspectral imagery

Q15. Why is weakly supervised learning useful in remote sensing?

It automates data augmentation
It reduces the model size
It trains using incomplete or coarsely labeled data
It requires spatial validation