AI for Maps, Remote Sensing, and Earth Observation

Ask what an image shows, where and when it was acquired, and what evidence a map-based claim still needs.

Two satellite images show a reservoir with visibly different shorelines. AI describes a dramatic decline and then supplies an exact percentage of water loss. The first statement may be supported by inspection. The percentage may require measurements and information the images do not contain.

Earth observation combines location, time, sensor behavior, processing, and interpretation. AI can help organize metadata, describe visible features, draft analysis code, and design validation. A credible result keeps each of those stages distinct.

The central question is not simply “What is in this picture?” It is “What can this particular data source establish about this place, at these times, at this resolution?”

Begin with a geographic question

Define the area of interest, date range, feature to measure, and intended use. “Study drought” is broad. “Compare mapped surface-water extent within a defined reservoir boundary between two specified acquisitions” is more manageable.

The narrower question still does not directly measure stored water volume, explain all causes of change, or describe present conditions. Those are additional questions requiring additional evidence.

Choose resolution according to the phenomenon. A large reservoir comparison and an individual drainage-ditch assessment need different spatial detail. A yearly comparison and a short-lived flood event need different timing.

Write down the smallest change you need to detect. Then ask whether the available data and uncertainty could support it. AI should not promise precision based on the number of decimal places in a generated table.

Inspect a real public comparison

NASA Earth Observatory published a Lake Mead comparison on July 22, 2022, with images dated July 6, 2000, and July 3, 2022. The page provides the comparison and credits the imagery to NASA Earth Observatory using USGS Landsat data. See “Lake Mead Keeps Dropping”.

Visual inspection of the published images shows a substantially narrower dark-water footprint in several areas in 2022, including the northern arm, and broader exposed light-colored margins. This is an observation of the displayed comparison, not a pixel-based area estimate or a measurement of storage volume.

The images describe the stated historical dates. They do not establish Lake Mead’s condition in September 2026. A current-condition question would require current records.

The published JPEGs are useful for interpretation and communication. They are not a substitute for obtaining the original analysis-ready imagery and its metadata when performing quantitative change detection.

Build a metadata record before analyzing pixels

For each scene, record the source, acquisition time, sensor and product identifier, processing level, coordinate reference system, pixel size, band definitions, quality information, and relevant scale factors or missing-value codes.

A coordinate reference system specifies how coordinates relate to locations. Two rasters can have the same number of rows and columns while representing different areas or different coordinate systems. Matching their array dimensions is not proof of geographic alignment.

The public Lake Mead article gives image dates and a visual context, but a quantitative project still needs the original scene identifiers and metadata. Do not reconstruct those identifiers from a JPEG filename or assume that every instrument detail can be inferred from the display caption. Resolve scene identity against the actual source catalog.

Use a table that makes the current evidence boundary explicit:

This record makes the exercise useful without pretending it has completed a remote-sensing analysis.

Understand spatial resolution and resampling

Pixel size describes the grid spacing associated with a product, but different bands and products can have different native resolutions. USGS describes Landsat 8 reflective bands at 30-meter resolution, a 15-meter panchromatic band, and thermal data acquired at 100 meters and resampled in products. Inspect the relevant product documentation rather than assigning one resolution to every band. See USGS’s Landsat 8 overview.

Resampling changes how data are represented on a grid. It does not create new independent ground detail merely by producing smaller pixels. A sharper-looking display and a more informative measurement are not necessarily the same thing.

Before subtracting or classifying images across dates, align the area, grid, resolution, and relevant processing assumptions. Choose a resampling approach appropriate to whether the data represents continuous values or categories, and document the choice.

Inspect geographic alignment using suitable stable features and metadata. Small offsets can create apparent change along boundaries even when the underlying feature has not moved.

Account for clouds, shadows, season, and missing observations

A dark region is not automatically water. Shadows and other surfaces can complicate interpretation. Clouds may obscure the feature entirely. A classification method needs to account for these conditions rather than treating every pixel as an equally reliable observation.

Use the appropriate quality information for the chosen product and document what was masked. If part of the area is missing on one date, do not silently count it as land or unchanged territory.

Timing also matters. Two observations from different seasons may reflect normal seasonal variation as well as longer-term change. Similar calendar dates can reduce one difference while leaving many others unresolved.

For the Lake Mead teaching pair, the dates are both in early July, but that alone does not establish identical illumination, atmospheric conditions, water management, or hydrologic context. A change-detection plan should list the remaining factors and the records needed to interpret them.

Choose a method that matches the evidence

A first stage can be visual annotation: identify visible water boundaries and mark uncertain areas. A more advanced stage might use a documented spectral index, supervised classification, or another established method with suitable bands and preprocessing.

Ask AI to explain what the proposed method measures and which inputs it needs. If it proposes an index requiring bands that are absent from the JPEG, the next step is to obtain appropriate data, not to calculate a substitute from display colors without justification.

Define thresholds using an appropriate development process. Keep validation observations separate from the examples used to choose the threshold. Record how ambiguous boundary pixels are handled.

A model-generated mask is a prediction. It becomes an evaluated map only after a suitable comparison with independent reference information. Even then, the evaluation applies to the stated data, region, and conditions.

Calculate area without pretending it is volume

Consider a separate fictional raster exercise with square, 30 m by 30 m pixels on a suitable area-measurement grid. Each pixel represents 900 m², or 0.09 hectares. If 100 whole pixels are classified as water, the simple mapped area is 90,000 m², or nine hectares.

That arithmetic does not account for classification error, mixed pixels, projection distortion outside the assumed suitable grid, or boundary uncertainty. It is not a measured area for Lake Mead.

Stored water volume requires additional information about the basin and water level or another appropriate volume-estimation method. A change in visible surface area cannot simply be renamed a percentage change in volume.

Similarly, a satellite observation of vegetation color does not by itself establish crop yield, ownership, habitat quality, or the cause of stress. Keep the measured or classified feature separate from the conclusion a user ultimately cares about.

Validate across relevant conditions

Plan an independent reference sample covering the conditions likely to cause errors: shoreline boundaries, shadows, narrow inlets, and clear examples of water and non-water, as appropriate to the project.

Do not reuse the same labeled locations to both tune the classifier and claim independent validation. Record the reference date and method so the comparison concerns the same phenomenon and period.

If you deliberately sample equal numbers from mapped water and mapped land, remember that the sample proportions may differ from the map’s area proportions. A simple overall average of sample correctness may not represent overall map accuracy. Report the sampling design and class-level results; use appropriate weighting and uncertainty methods for area-wide estimates.

This is a place where specialist review can materially change the conclusion. AI can prepare the records and calculations, but an attractive accuracy percentage cannot repair an unsuitable sampling design.

Publish a map with the information needed to interpret it

A map should identify the place, acquisition dates, data source, legend, relevant units, processing method, and limitations. Include scale and orientation where they help the intended reader, and use an appropriate coordinate reference for quantitative work.

Make missing or uncertain regions visible. If two dates have different coverage, explain the comparison boundary. Avoid a title that states a causal conclusion the analysis did not test.

A suitable title for our initial exercise is “Visual comparison of Lake Mead imagery, July 2000 and July 2022.” “Exact water-volume loss caused by a single factor” would claim substantially more than the exercise establishes.

Keep the analysis version linked to the map. If a classification threshold or mask changes, regenerate the reported area and update the method note rather than editing the number manually.

A reusable prompt

Inspect these imagery metadata and the research question. Identify location, date, sensor, product, coordinate system, grid, resolution, band, quality-mask, and alignment issues. Separate visible observations from measured quantities and causal interpretations. Propose a change-detection and independent validation plan using the available data, and identify data still required. Do not infer current conditions from historical imagery, assign precise area from an uncalibrated display, or equate surface area with storage volume.

For students: read the image and the caption together

Open the NASA comparison and write three observations visible in the images. Then write three questions the images alone cannot answer. Keep these lists separate.

Geography students can focus on location, dates, and map communication. Science students can identify additional measurements needed to investigate causes. Computing students can design a metadata checker or a synthetic raster calculation without claiming to have analyzed the reservoir.

For advanced GIS work, use authorized public scene data and document the actual processing steps. Submit the methods and uncertainty record alongside the final map so the assignment assesses analysis rather than visual polish alone.

Practice: draft a defensible change-detection plan

Use the Lake Mead pair to define a historical surface-water question. List the original data and metadata you would obtain, how you would align and screen the scenes, the feature you would classify, and the independent reference information you would seek.

Add the fictional 100-pixel area calculation as a separate unit exercise. Explain why it does not provide a reservoir measurement. Draft a map caption that identifies dates and limits without attributing a cause or claiming current conditions.

Completion check: The place and historical dates are correct; observations are separated from unperformed measurements; missing metadata is listed; 100 square 30-meter pixels equal nine hectares in the stated exercise; and the plan includes independent validation and an explicit distinction between surface area and water volume.

Stretch: Design a stratified reference sample covering clear interiors, boundaries, and difficult conditions. Explain which population each reported accuracy measure describes and what weighting would be needed for an area-wide estimate.