Estimating Biomass in Trans-Himalayan Rangelands Using Sentinel-2 Data and Machine Learning Approaches

The Great Grass Detective: A Story of Mountains, Satellites, and Smart Computers

Imagine the Earth's skin is covered in vast, wild carpets of grass and shrubs called rangelands. These aren't just pretty scenery; they are the giant dining tables for millions of cows, sheep, and goats, and they help keep the planet's water clean and its air fresh. But here's the problem: these grassy carpets are getting worn out. Too many animals eating at once, changing weather, and human pressure are turning lush meadows into dusty patches. To fix this, we need to know exactly how much grass is left, but counting it the old-fashioned way is a nightmare. It involves hiking up steep, rocky mountains, cutting tiny squares of grass with scissors, drying them in the sun, and weighing them. It's slow, expensive, and you can't do it everywhere at once.

This is where the magic of "remote sensing" and "machine learning" comes in. Think of remote sensing as a giant, high-tech camera in the sky (a satellite) that can see colors our eyes can't, like the invisible "red edge" of light that plants glow with when they are healthy. Machine learning is like a super-smart computer student that can look at millions of these satellite photos and learn to guess how much grass is on the ground without ever stepping foot on it. Scientists have been trying to teach these computers to be accurate, but mountainous areas are tricky because the shadows and steep slopes confuse the cameras. The big question is: Can we build a smart system that can accurately "weigh" the grass in these tough, high-altitude mountains just by looking at the sky?

---

The Paper's Mission: Teaching Computers to Count Grass in the Clouds

This research paper is like a field guide for a new kind of detective work. The authors, a team of scientists from Pakistan and Canada, set out to solve the mystery of how much grass (biomass) is growing in the Trans-Himalayan rangelands of Upper Dir and Swat. These are rugged, high-mountain areas where the terrain is so steep and the weather so wild that traditional counting methods are nearly impossible. Instead of sending teams of hikers to clip grass, the team decided to use data from the Sentinel-2 satellite, which takes incredibly detailed pictures of the Earth every few days.

They treated the satellite images like a giant puzzle. First, they went into the field during the summer of 2022 and 2023 to collect "ground truth." They hiked to random spots, measured the actual dry weight of the grass in small squares, and recorded the GPS coordinates. This was their answer key. Then, they looked at the satellite photos of those exact same spots to see what the satellite "saw." They tested 50 different mathematical formulas (called vegetation indices) that try to turn light into numbers. It was like trying 50 different recipes to see which one made the best cake.

The Discovery: The "Red Edge" Wins the Race

The team found that not all recipes were created equal. Some of the old, standard formulas, like the famous NDVI (which measures how green the plants look), were okay, but they missed the finer details. The real star of the show turned out to be a formula called REDVI1. This formula uses a special part of the light spectrum called the "Red Edge," which is like a secret handshake between the satellite and the plant's chlorophyll.

When they tested their theories, REDVI1 was the clear winner for prediction accuracy in the initial screening. In 2023, it predicted the amount of grass with an accuracy score (R²) of 0.86, which is a very strong match. The standard green-measuring formula, NDVI, only scored 0.7376. It's as if REDVI1 could see the health of the grass from a mile away, while NDVI was squinting and guessing.

But knowing the best formula wasn't enough; they also needed the smartest computer brain to do the math. They tested three different machine learning "students":

1. PLSR (a linear thinker).
2. SVR (a support vector machine).
3. RFR (Random Forest Regression).

The Random Forest (RFR) model was the champion. It didn't just guess; it built a forest of decision trees that could handle the messy, complicated reality of the mountains. It was much more stable and accurate than the others, with a performance score (NSE) of 0.92 in 2023. The paper explicitly notes that the older, simpler linear methods and the other machine learning models struggled more with the weird shadows and steep slopes of the mountains. The RFR model was the only one that could consistently ignore the noise and find the signal.

The Map: Where the Grass is Fat and Where it is Thin

Using their winning machine learning model (RFR) to validate the data, the team then created a detailed map of the entire region. However, for the actual final map, they didn't use the complex computer model directly. Instead, they applied a simpler linear regression equation based on REDVI2. While REDVI1 was the top predictor for accuracy in the initial screening, the team found that REDVI2 had the highest correlation for the specific mapping phase (with an R² of 0.7204 in 2023), so they used that formula to calculate the grass weight for every single pixel on the map.

The results revealed a fascinating story about the land:

• The Average: Across the whole study area, the average amount of dry grass was 5,896.78 kg/ha.
• The Extremes: The grass wasn't spread out evenly. Some spots had almost no grass (near towns and rivers where animals graze all year), while other spots were incredibly lush. The highest patches had up to 11,516.7 kg/ha of grass.
• The Pattern: The paper found that the grass near villages and water sources was often degraded, with low biomass (less than 3,500 kg/ha). This is because the animals have been eating there for too long without a break. However, the high mountains, which are covered in snow for most of the year, had thick, healthy grass (over 7,000 kg/ha) because the snow kept the animals away, giving the plants a chance to grow back.

What This Means for the Future

The authors are careful not to claim they have "solved" the problem of rangeland management forever. Instead, they suggest that this method offers a powerful new tool. They showed that you don't need to cut grass to know how much is there; you can use a satellite and a smart computer to get a very accurate picture.

This is a big deal for the people living there. If local leaders can see a map that says, "Hey, this valley is running out of grass," they can move the animals to a different area before the land gets destroyed. It turns guesswork into science. The paper concludes that this approach is accurate, scalable, and non-invasive, offering a way to protect these fragile mountain ecosystems and the livelihoods of the people who depend on them. It's a reminder that sometimes, the best way to understand the ground is to look at it from the sky.