Reflectance Multispectral Imaging for Soil Composition Estimation and USDA Texture Classification

This paper presents a cost-effective, field-deployable multispectral imaging system combined with machine learning that achieves high accuracy (R² up to 0.99 and over 99% classification accuracy) in predicting soil composition and USDA texture classes, offering a rapid, non-destructive alternative to traditional laboratory testing for agriculture and geotechnical engineering.

The Gist

Imagine you are a farmer or a civil engineer trying to figure out what your soil is made of. Is it mostly sand (like a beach)? Is it mostly clay (like a potter's mud)? Or is it a mix of silt (like flour)?

Traditionally, to find this out, you have to take a dirt sample, pack it in a box, mail it to a lab, and wait days or weeks for a scientist to sift it through screens and measure it with chemicals. It's slow, expensive, and requires a lot of manual labor.

This paper introduces a high-tech, instant "soil scanner" that does the same job in seconds, right in the field, using a camera and a bit of computer magic.

Here is the breakdown of how they did it, using some simple analogies:

1. The "Super-Eye" Camera (Multispectral Imaging)

Normal cameras (like on your phone) see the world in three colors: Red, Green, and Blue. They are like a person who only knows three shades of gray.

The researchers built a custom camera that sees 13 different "colors" of light, stretching from deep violet (which humans can't see) to near-infrared (which is just beyond red).

• The Analogy: Imagine you are trying to identify different types of fruit. A normal camera just sees "round and red." But this special camera can see the specific way light bounces off an apple versus a cherry versus a tomato.
• How it works: They shine 13 different colored LED lights on the dirt. The dirt reflects these lights back to the camera. Because sand, clay, and silt are made of different minerals, they reflect these 13 lights in unique patterns. It's like every type of soil has a unique "fingerprint" made of light.

2. The "Recipe" (The Soil Mix)

To teach the computer how to read these fingerprints, they needed a "training manual."

• They went to three different places in Sri Lanka to find pure samples: one place with mostly clay, one with mostly silt, and one with mostly sand.
• They then mixed these three "pure ingredients" together in a kitchen, creating hundreds of different "recipes" (mixtures) to cover every possible soil type defined by the USDA (the US Department of Agriculture).
• They measured these mixtures in a lab first to know the exact truth (the "ground truth"), so the computer could learn to match the light patterns to the correct recipe.

3. The "Brain" (Machine Learning)

Once they had the photos and the lab data, they fed it into a computer brain (Machine Learning). They tried three different ways to solve the puzzle:

• Strategy A: The Direct Guess (The "Gut Feeling")
  The computer looks at the light pattern and immediately says, "This is a 'Sandy Clay Loam'!" It skips the math and goes straight to the answer.

  • Result: This was the fastest and most accurate method, getting it right 99.5% of the time.

• Strategy B: The Recipe Calculator (The "Chef")
  The computer looks at the light pattern and calculates the exact percentages: "This is 40% sand, 30% clay, and 30% silt."

  • Result: This was incredibly precise, almost perfect (99.9% accuracy in predicting the percentages).

• Strategy C: The Map Reader (The "Indirect" Way)
  The computer uses the "Chef" method (Strategy B) to get the percentages, and then draws a dot on a standard USDA triangle map to see which zone that dot falls into.

  • Result: This was also very good (97% accuracy), but slightly less accurate than the direct guess. Why? Because if the computer is off by just a tiny fraction on the percentage, the dot might land on the wrong side of a line on the map, changing the whole classification.

4. Why This Matters

The paper proves that you don't need a million-dollar lab or a satellite to know your soil.

• Speed: It takes seconds, not weeks.
• Cost: The camera they built is cheap and portable.
• Usefulness:
  • Farmers can instantly know if they need to add water or fertilizer.
  • Engineers can quickly check if the ground is stable enough to build a house on (clay swells when wet and can crack foundations; sand drains too fast).
  • Environmentalists can monitor soil health without digging up the earth.

The Bottom Line

Think of this system as a "Soil Translator." It translates the invisible language of light bouncing off dirt into a simple, easy-to-understand recipe or a soil type name. It turns a complex, slow scientific process into something as easy as taking a photo with a specialized camera.

The researchers found that while the "indirect" method (calculating percentages first) is great for understanding what the soil is made of, the "direct" method (just guessing the type) is the most reliable for getting the right answer quickly. Either way, they have built a tool that could revolutionize how we manage our land.

Technical Summary

1. Problem Statement

Soil texture (the relative proportions of clay, silt, and sand) is a critical parameter for agriculture, geotechnical engineering, and environmental monitoring. It governs water retention, erosion risk, and load-bearing capacity. However, traditional determination methods rely on laboratory-based particle size analysis (sieve and hydrometer methods), which are:

• Time-consuming and labor-intensive: Often taking 1–2 days per sample.
• Non-portable: Require specialized equipment and controlled environments.
• Destructive: Often require soil removal and processing.

Existing sensing alternatives have limitations:

• RGB Imaging: Low cost but highly sensitive to lighting and surface conditions; cannot estimate continuous composition fractions.
• Hyperspectral Imaging (HSI): High accuracy but expensive, bulky, and generates massive data volumes unsuitable for routine field deployment.
• Satellite Remote Sensing: Limited by coarse spatial resolution and the requirement for bare soil pixels.

The paper addresses the need for a cost-effective, portable, and field-deployable system that can non-destructively estimate both soil composition (clay, silt, sand percentages) and USDA texture classes with high accuracy.

2. Methodology

A. Hardware System: Custom Multispectral Imaging (MSI)

The authors developed an in-house, low-cost MSI system designed for field use:

• Sensor: FLIR BFS-U3-13Y3M monochrome machine vision camera (1.3 MP).
• Illumination: Thirteen narrowband LED clusters covering the 365 nm to 940 nm spectrum (UV to NIR).
• Wavelengths: 365, 405, 473, 530, 575, 621, 660, 735, 770, 830, 850, 890, and 940 nm.
• Setup: Enclosed in a custom dark chamber with matte black interiors to eliminate ambient light interference.
• Operation: Controlled by an Arduino Due for LED switching and image acquisition.

B. Dataset Preparation

• Source Materials: Three "endmember" soils were collected from Sri Lanka (Menikhinna for clay, Gelioya for silt, Chavakachcheri for sand).
• Ground Truth: Source soils were analyzed via standard sieve and hydrometer methods to determine exact clay/silt/sand percentages.
• Mixtures: 22 unique mixture ratios were created to cover all 12 USDA texture classes, resulting in 440 specimens for training/testing and 84 specimens for external validation.
• Total Data: 524 specimens, generating a dataset of 52,400 block-level observations (100 blocks per specimen).

C. Image Processing Pipeline

1. Preprocessing:
   • Dark Current Correction: Subtraction of dark frames to remove sensor noise.
   • ROI Selection: Cropping a fixed 100×100 pixel region of interest.
   • Contrast Normalization: Application of a bounded hyperbolic tangent ($\tanh$) mapping to normalize intensities and suppress outliers while preserving mid-range variations.
2. Feature Extraction:
   • The 100×100 ROI was divided into a 10×10 grid of non-overlapping blocks.
   • Mean intensity was calculated for each block across all 13 bands, creating a 100 × 13 feature matrix per sample.
3. Dimensionality Reduction:
   • Linear Discriminant Analysis (LDA) was applied to reduce the 13 spectral bands to 5 discriminative components, maximizing class separability while minimizing intra-class variance.

D. Modeling Strategies

The study evaluated three distinct approaches using the LDA-reduced features:

1. Direct Classification: A supervised classifier maps spectral features directly to the 12 USDA texture classes.
2. Composition Regression: Regression models predict the continuous percentages of Clay, Silt, and Sand.
3. Indirect Classification: The predicted composition percentages from (2) are mapped to USDA texture classes using the standard USDA soil texture triangle decision rules.

Algorithms Tested: K-Nearest Neighbors (KNN), Random Forest (RF), Decision Trees (DT), CatBoost (CB), and XGBoost (XGB).

3. Key Contributions

• Hardware Innovation: Development of a low-cost, 13-band MSI system specifically optimized for soil texture characterization, bridging the gap between cheap RGB cameras and expensive HSI systems.
• Dual-Output Framework: A unified pipeline that supports both direct classification (for rapid screening) and composition regression (for detailed agronomic analysis).
• Comprehensive Evaluation: A rigorous comparison of direct vs. indirect classification strategies, including an external validation set to test generalization.
• Curated Dataset: Creation of a ground-truthed dataset of synthetic soil mixtures covering the full USDA texture triangle.

4. Key Results

A. Direct Classification Performance

• Best Model: K-Nearest Neighbors (KNN).
• Accuracy: 99.55% (Mean) with a standard deviation of 0.0007.
• Metrics: Macro F1-score of 0.9960 and Recall of 0.9966.
• Observation: All models performed well, but KNN showed the highest stability. Confusion was minimal, occurring mostly between adjacent classes (e.g., Silt vs. Silt Loam).

B. Composition Regression Performance

• Best Model: KNN.
• Performance (Testing Set):
  • Clay: $R^2 = 0.9993$, RMSE = 0.53%.
  • Silt: $R^2 = 0.9988$, RMSE = 0.92%.
  • Sand: $R^2 = 0.9982$, RMSE = 1.17%.
• External Validation: The model maintained high accuracy on unseen data ($R^2 > 0.98$ for all components), demonstrating strong generalization.
• Note: Sand estimation was slightly more challenging (higher RMSE) than clay or silt, likely due to quartz dominance in sand having fewer distinct absorption features in the 365–940 nm range.

C. Indirect Classification Performance

• Best Model: KNN.
• Accuracy: 96.98%.
• Comparison: While high, this is approximately 2.57% lower than the direct classification approach.
• Reason for Drop: The indirect method suffers from error amplification near the boundaries of the USDA texture triangle. Small errors in predicted percentages can cause a sample to cross a decision boundary, resulting in a different class label.

5. Significance and Conclusion

• Field Deployability: The proposed system proves that compact, low-cost MSI devices can replace slow laboratory tests for routine soil screening in agriculture and geotechnical engineering.
• Operational Trade-offs:
  • Direct Classification is superior for rapid, high-accuracy decision-making (e.g., "Is this soil suitable for foundation construction?").
  • Indirect/Regression provides interpretable data (exact % of clay/silt/sand) essential for precision agriculture (fertilizer/irrigation planning), despite a slight accuracy penalty.
• Scientific Validation: The results confirm that soil texture classes form distinct, separable clusters in the multispectral feature space, validating the physical basis of the USDA texture triangle when viewed through optical reflectance.
• Impact: This work offers a scalable solution for soil characterization, enabling non-destructive, real-time monitoring of soil health and stability, which is critical for preventing infrastructure failures and optimizing agricultural yields.