Two pixels have identical spectral signatures in a Landsat image -- both show short green vegetation. One is a golf course; the other is a wheat field. What type of additional information is needed to distinguish them?
AHigher radiometric resolution
BContextual and ancillary information: the golf course is surrounded by urban development and has distinctive spatial patterns (fairways, greens), while the wheat field is in a rural area with regular rectangular parcels
CThermal infrared data to measure their temperature difference
DHyperspectral data to identify grass species
This illustrates the land cover vs. land use distinction. Both have the same land cover (managed grass) but different land uses (recreation vs. agriculture). Spectral signatures alone cannot distinguish them. Context -- surrounding land cover, parcel shape, proximity to urban areas, cadastral boundaries -- is required. This is why LULC classification increasingly incorporates spatial context, ancillary GIS data, and object-based approaches rather than relying on pixel spectra alone.
Question 2 True / False
A single satellite image is sufficient to produce an accurate land cover map for a region with diverse agricultural crops.
TTrue
FFalse
Answer: False
Different crops may be spectrally identical at a single date but have different phenological calendars -- planting, growth, and harvest occur at different times. A multi-temporal approach using images across the growing season captures these phenological differences, dramatically improving crop type discrimination. Winter wheat is green in March when corn fields are bare soil; by July the pattern reverses. Time-series classification exploiting these phenological signatures is now standard for agricultural LULC mapping.
Question 3 Short Answer
Explain the fundamental difference between pixel-based LULC classification and the approach used by Google's Dynamic World product.
Think about your answer, then reveal below.
Model answer: Traditional pixel-based classification assigns each pixel a single discrete class label (forest, water, urban) based on spectral values at one or a few dates. Dynamic World uses a deep learning model (neural network) trained on billions of labeled pixels to produce per-pixel probability estimates for each class at every Sentinel-2 observation (every 2-5 days). This yields continuous probability surfaces rather than hard classifications, allows users to choose their own confidence thresholds, and provides near-real-time updates. The temporal density enables tracking of land cover transitions as they happen rather than in annual snapshots.
Dynamic World represents a paradigm shift: from periodic hard classification to continuous probabilistic monitoring, enabled by cloud computing and deep learning at planetary scale.