Graduate school occasionally surprises you.

You sign up for a GIS class expecting maps, coordinates, and perhaps the occasional argument with software that behaves like it personally dislikes you. Then suddenly, you are staring at satellite imagery from Peru, comparing a river before and after a flood, quietly realizing that modern geography has evolved into something resembling detective work from orbit.

Not bad for pixels.

For one of my assignments in REM-617: Image Analysis & Information Extraction, I worked on change detection analysis using Sentinel-2 satellite imagery and ENVI software to examine flood impacts along the Mala River in Peru after a major El Niño event in 2017. The assignment focused on comparing pre-flood and post-flood imagery to identify where environmental changes occurred and how severe they were.

And yes, this is apparently my life now.

Watching floods from space.

The River Before the Problem

The lab used two Sentinel-2 images. One image captured the Mala River area before flooding in February 2017. The second image showed conditions after severe flooding in March 2017 caused by unusually intense rainfall linked to El Niño. Heavy rains west of the Andes triggered flooding, landslides, and extensive property damage in coastal communities near Lima, Peru. The Mala River overflowed its banks and became a useful case study because cloud-free imagery happened to be available before and after the event. Convenient for science. Less convenient for the people living there.

At first glance, the exercise felt deceptively simple.

Put the images side by side.

Zoom in.

Look for differences.

Easy.

Except it turns out rivers are messy creatures. Flooding does not politely stay inside obvious lines like a diagram in a textbook. Water spreads. Sediment moves. Agricultural land changes color. Moisture alters reflectance. Nature, as usual, refuses to cooperate with human expectations.

Geography, But Slightly More Suspicious

One thing I appreciated about this assignment was how visual the process became.

The software allowed the before-and-after images to be geographically linked. Zoom into one image and the other follows perfectly, creating a synchronized comparison of the same landscape before and after the flood. Suddenly, you are not just looking at maps. You are investigating evidence.

This part felt oddly satisfying.

Like being a forensic investigator, except your witness is a satellite orbiting hundreds of kilometers above Earth and your suspect is a river that decided rules were optional.

The flood extent became immediately visible. River channels expanded. Surface characteristics changed. Areas near the river corridor displayed noticeable differences in texture and coloration, hinting at sediment redistribution and newly saturated ground.

Enter PCA: The Part Where the Computer Sees What You Missed

Then came Principal Component Analysis (PCA).

Now, PCA sounds intimidating, mostly because academia occasionally enjoys naming things in ways guaranteed to frighten students.

In reality, PCA is simply a method for reducing complexity and highlighting important differences in data. Think of it as asking the computer:

“Please stop showing me noise and show me what actually changed.”

The PCA transformation reduced spectral redundancy and emphasized variations that were difficult to see in standard imagery. Instead of relying solely on ordinary RGB images, different principal component bands amplified subtle changes across vegetation, soil moisture, and river features. Bright contrasts in the imagery pointed to places where something significant happened. And something clearly had.

What struck me was how much environmental information was hiding in plain sight.

To the naked eye, some areas looked ordinary.

To spectral analysis?

Not even close.

The Principal Component Analysis (PCA) was applied to the MalaRiver_2017-02-20 dataset to reduce spectral redundancy and highlight variance in the imagery. The resulting PCA bands were visualized using a composite of PC Band 4 (Red), PC Band 3 (Green), and PC Band 2 (Blue).

PC Band 1 contained overall brightness information and was not used for change detection. Instead, PC Bands 2 and 3 revealed significant spectral variation across the landscape.

The PCA composite enhanced differences between land cover types, including vegetation, bare soil, and river features. Areas of potential change appear as bright color contrasts, particularly in the highlighted region.
The PCA output revealed distinct spectral differences that were not easily visible in the original RGB imagery. The river channel and surrounding areas exhibited strong color variations, indicating possible changes in sediment, moisture, or vegetation.

The boxed region showed enhanced contrast in the PCA image compared to the original, suggesting localized changes in land cover or surface conditions. PCA effectively amplified subtle variations, making it a useful method for detecting change in multispectral imagery.

The computer quietly raised an eyebrow and said:

“You may want to look over here.”

The eigenvalue plot shows a sharp decline after the first principal component, indicating that most of the variance in the dataset is captured by the first few components. PC1 contains the majority of the image brightness information, while PC2 and PC3 capture meaningful spectral variation.

The remaining components (PC4 and beyond) contribute minimal additional information and are likely dominated by noise. This confirms that PCA effectively reduces dimensionality while preserving the most relevant features for analysis and change detection.

Why Moisture Matters More Than Pretty Pictures

The most useful part of the analysis came from using SWIR Band 11.

Shortwave Infrared, or SWIR, is particularly sensitive to moisture and water content. Which, when studying floods, is exactly what you want.

This is where satellite science becomes unexpectedly clever.

Floodwater, wet soil, and sediment-laden surfaces reflect energy differently than dry land. By subtracting pre-flood and post-flood SWIR imagery, the analysis highlighted areas where moisture conditions changed significantly after the flood event. Bright and dark contrasts revealed the story of where water arrived, lingered, or altered the landscape.

Translation into ordinary human language:

The satellite basically said:

“Yep. This place got soaked.”

Repeatedly.

The preview/difference image shows pixel-level changes between the two dates. Positive values (brighter tones) indicate increased reflectance, while negative values (darker tones) indicate decreased reflectance. These changes correspond to environmental differences observed after the flood event.

The difference image generated using SWIR Band 11 highlights areas of change between the pre-flood and post-flood images. Bright areas represent increased reflectance, which may indicate drying surfaces or exposed sediment, while dark areas represent decreased reflectance, likely associated with increased moisture or water presence.

Significant changes are observed along the river corridor, where strong contrast indicates flooding impact and sediment redistribution. In contrast, the selected area shows minimal variation, suggesting limited change during the study period.

The use of SWIR Band 11 is appropriate because it is sensitive to moisture and surface conditions, making it effective for detecting flood-related changes.

Turning Chaos into Evidence

The workflow eventually moved into thresholding, smoothing, and vectorization.

Which sounds very technical and slightly unpleasant.

But conceptually, it is simple.

Thresholding turns a complicated image into a decision map.

Essentially:

Thresholding is the process of turning a complex image into a simple “Yes/No” map. You are telling ENVI: “If the change was large enough to be a flood, color it red; if not, ignore it.” This converts your data into a Classification Image.

Flood or not flood.

Enough change? Keep it.

Too little change? Ignore it.

This is purely a symbology step. It doesn’t change the pixels, but it attaches labels and colors to the data values. By naming the changed pixels “Flood Impact,” you ensure that any future maps, legends, or vector exports (like the Shapefile in Task 5) will be properly labeled.

Then the software converts pixels into polygons, creating professional flood boundaries that can actually be used for planning, emergency management, or analysis. What begins as spectral noise slowly becomes something understandable.

A map.

This process is called Vectorization. It converts your raster data (individual red pixels) into vector polygons (smooth outlines). This is critical for professional mapping because vectors can be scaled without losing quality and are the standard format for sharing flood data with emergency services or government agencies.

Evidence.

A measurable impact.

In this case, the estimated flood extent reached approximately 4.76 square kilometers, concentrated primarily along the river corridor and nearby agricultural areas. That estimate came directly from analyzing changes in moisture-sensitive spectral bands after the flood event.

This table is our Attribute Database. Every shape you created in the vectorization step is linked to a row in this table. By adding the area column, you have performed a Geometric Intersection. The software multiplies the total number of pixels identified as “Flood” by the ground area each pixel represents (10m x 10m = 100m²). This provides the definitive, quantifiable evidence of the flood’s extent for your report.

That is the thing about GIS people rarely appreciate.

Maps are not merely pictures.

Good GIS answers questions.

Sometimes uncomfortable ones.

How much land flooded?

Where did it spread?

Who was affected?

What changed?

And perhaps most importantly:

Can we prove it?

The flood extent was estimated at 4.76 km² based on Sentinel-2 band difference analysis. Using SWIR Band 11, which is sensitive to moisture and water content, the method accurately identified areas of increased water presence after the flood event. The detected changes are concentrated along the river corridor and nearby agricultural fields, consistent with expected flood patterns in the study area.

The Strange Joy of Watching Geography Misbehave

I will admit something.

There is something deeply satisfying about this kind of work.

Not the flood itself, obviously.

But the ability to reconstruct an event using nothing but imagery, algorithms, and careful analysis.

You begin with two dates.

A river.

A few satellite bands.

And somehow end up telling a story about environmental change from space.

Which is objectively cool, even if GIS people occasionally struggle to explain this without sounding like they are describing spreadsheets with better marketing.

Graduate school has me doing strange things lately.

One week I am reading about geospatial theory.

The next week I am staring at Peru from orbit thinking:

“Well, that river certainly had a rough month.”

And somehow, weirdly enough, I enjoy it.