Gains:
- Ability to interpret GIS layers, NDVI and remote sensing band/resolution concepts
- Index interpretation, change analysis and map narrative drafting with AI
- Ability to verify AI's satellite/map interpretation with ground truth and metadata
You are an environmental engineer working on a wetland conservation project. The council wants you to investigate complaints that the lakeshore reed bed has shrunk over the last five years. Before going to the field, you need to make a preliminary evaluation at your desk: you have satellite images from two different years, a GIS (Geographic Information System) project and a language model. You can have LLM interpret the NDVI summary of the images and write a draft of the change narrative. However, a satellite pixel represents a 10-meter square, and you can only understand whether there is a cloud shadow, drying reed or seasonal water level in that square with ground truth. In this unit, we cover the concepts of remote sensing and GIS, index/change interpretation with AI, and most importantly, the limits of this interpretation.
GIS layer logic and remote sensing fundamentals
GIS organizes geographic data into overlapping layers: each layer carries a theme (land use, water bodies, elevation, roads) and is aligned in a common coordinate system. An analysis starts with overlapping the correct layers in the correct projection; the wrong coordinate system systematically biases areas and distances.
When interpreting a remote sensing image you must distinguish three types of resolution:
Resolution type
What does it tell?
example
limitation
spatial
Size of a pixel on the ground
10m, 30m
Small objects remain sub-pixel
spectral
How many and which wavelength band?
Red, NIR, SWIR
Less band = less discrimination power
temporal
How often the same place is viewed
5 days, 16 days
Data loss on cloudy days
Radiometric
Luminance sensitivity (bit depth)
8-bit, 12-bit
Low depth = poor discrimination
NDVI: formula and interpretation
The most common index measuring the vitality of vegetation is NDVI (Normalized Difference Vegetation Index). A healthy plant strongly reflects near infrared (NIR) and absorbs red (RED):
NDVI = (NIR − RED) / (NIR + RED)Value range: −1 ... +1 < 0 : water, snow, cloud 0.0 - 0.2 : bare soil, rock, built area 0.2 - 0.5 : sparse/herbaceous vegetation, stressed cover 0.5 - 0.9 : dense, healthy green cover
# NDVI of a small sample block of pixels import numpy as npnir = np.array([[0.42, 0.38], [0.11, 0.45]]) # near infrared reflectancered = np.array([[0.06, 0.09], [0.10, 0.05]]) # red reflectancendvi = (nir - red) / (nir + red)print(np.round(ndvi, 2))# [[0.75 0.62]# [0.05 0.80]]# Bottom-left pixel (0.05): not a plant — could be bare ground or water/shadow
Tip: NDVI is not an absolute measure of "health" but a condition-dependent indicator. When comparing two dated NDVIs of the same area, make sure the images were taken in similar seasons, similar sun angles, and cloud-free conditions. Comparing the appearance of late summer to spring produces an unreal “plant loss.”
Change analysis (change detection)
The simple way to find the difference between two dates is to take the NDVI difference, but this difference is only meaningful if the images can be compared.
ΔNDVI = NDVI_2025 − NDVI_2020 ΔNDVI << 0 : cover loss possible (drying, cutting, construction) ΔNDVI ≈ 0 : no change ΔNDVI >> 0 : cover increase possible (greening, restoration) ATTENTION: The difference may reflect actual change as well as season/cloud/sun angle/calibration differences. Difference = real change + noise.
Caution: Cloud, cloud shadow and seasonality are the biggest detractors of change analysis. The area under a cloud shadow gives low NDVI and looks like "vegetation loss". When LLM performs image interpretation, the model may mistake these artifacts for the real event. Be sure to check the image metadata (date, cloud percentage, sun angle).
Index interpretation and map narrative with AI
LLM is useful in translating numerical index summaries into plain narrative and producing draft reports. Give it the raw NDVI statistics and metadata and ask it to generate the comment for you to verify.
WEAK PROMPT: "Has the wetland become smaller in this satellite image?" STRONG PROMPT: "I give two dated NDVI statistics (same area, same sensor): 2020-07-15 avg NDVI=0.61 cloud=3% sun angle=58°2025-07-12 avg NDVI=0.42 cloud=22% sun angle=55°Area of interest: lakeshore reed bed, ~40 hectares.Your task:1) Calculate ΔNDVI and list possible interpretations.2) List what reasons (cloud, shade, water level, season, calibration) this decline could be due to (cloud, shade, water level, season, calibration) OTHER THAN ACTUAL cover loss.3) For each possible cause, suggest a VERIFICATION step to distinguish it (e.g. cloud-free image selection, field visit, water level recording).4) Accurate claim 'reedbed has shrunk by X%' DO NOT SETUP; state that this can only be said if the ground truth confirms it.”
The strong prompt directs the model not to a single definitive conclusion but to a plan of alternative explanations and verification. The fact that the 2025 image contains 22% clouds makes the decline suspicious in itself.
Ground truth verification
The gold standard of remote sensing interpretation is verification points with observations and measurements made in the field. You record the actual land cover with GPS at several reference points and compare it with the image classification. If the field says "dry reeds" in the pixel where the AI or algorithm says "water", the classification is unreliable.
mini case
In environmental impact monitoring of a mining site, the team drafted “15 hectares of forest loss due to mining activity” based on the NDVI reduction calculated from the satellite image and had the interpretation expanded into an LLM; the model produced a fluid, persuasive narrative. However, when the report was checked by the team on the field, it was seen that part of the area with low NDVI was actually a harvested agricultural parcel and part of it was a cloud shadow that coincided with the day of the image shooting. The actual forest loss was about 4 hectares, not 15. If the metadata (cloud mask) and land use layer had been checked beforehand, the error would have been prevented in the first place. LLM's neat sentences did not make wrong input right.
Common mistakes
- Comparing different season/sun angle images and mistaking the seasonal difference for "change".
- Interpreting cloud and cloud shadow as actual cover loss; bypassing cloud mask.
- Making accurate judgments from pixel data about small objects exceeding spatial resolution.
- Miscalculating areas and distances due to coordinate system/projection mismatch.
- Thinking that NDVI is an absolute measure of health and treating water and shadow pixels as if they were plants.
- Putting LLM's fluid map narrative into the report as a final finding without ground truth.
- Reporting results without documenting metadata (date, sensor, cloud %, processing level).
In summary
- GIS analyzes are built on layers aligned in a common coordinate system; Projection error distorts the result.
- Spatial, spectral, temporal and radiometric resolution limit what an image can show.
- NDVI = (NIR−RED)/(NIR+RED) is a conditional indicator of plant vigor, not an absolute measure of health.
- In change analysis, difference = true change + noise; Cloud, shadow and season are the biggest deceivers.
- The LLM produces the index interpretation and map narrative outline, but does not "see" the satellite image; Comments according to the number you give.
- AI interpretation and automatic classification cannot be translated into final environmental findings without verification with ground truth and metadata.
Application task
Create or assume two dated NDVI summaries for a field (either of your choice or a sample); Note the metadata (date, cloud percentage, sun angle, sensor) for each date. Calculate ΔNDVI. Write a strong prompt and have the LLM ask not just for the actual cover change, but for alternative explanations (cloud, shadow, season, water level, calibration) and a verification step for each. Then draft a “ground truth verification plan”: at least three reference points, what will be measured in the field, and how it will compare with the image interpretation. Finally, mark which statement in the narrative produced by LLM is "cannot be included in the report without verification" and write the reason.