Reflectance Multispectral Imaging for Soil Composition Estimation and USDA Texture Classification

arXiv:2602.22829 · cs.CV, eess.SP · Submitted 2026-02-26 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: Today's paper: "Reflectance Multispectral Imaging for Soil Composition Estimation and USDA Texture Classification".

Jane: This manuscript proposes a robust and field deployable multispectral imaging (MSI) system combined with machine learning to accurately predict soil composition and the United States Department of Agriculture (USDA) texture…

Tom: First, who's behind it and why it matters.

Title and authors: Jane: So, to kick things off, we’ve got a paper that basically takes what used to take weeks in a lab and tries to figure out soil texture just by looking at its reflection with a camera.

Tom: That's right, Jane; essentially, they developed this multispectral imaging system combined with machine learning to predict how different USDA soil textures look based on their spectral signatures. It’s a pretty neat way to get detailed soil information without having to physically dig up and test every single sample manually.

Lu: What really stands out from the summary is how they built a whole pipeline, starting from designing this custom MSI device that captures thirteen spectral bands, right down to using Linear Discriminant Analysis to make those features even cleaner before feeding them into their various AI models.

Meng: From my side, the summary mentions they tested three different ways of doing things—direct classification, estimating the clay and sand percentages through regression, and then using a rule-based mapping system for an indirect texture classification. That shows a really thorough approach to solving this problem.

Lalam: What I find most impactful is how they show that their feature space is actually very good at separating these different soil types, especially when using the K-Nearest Neighbors model, which got nearly perfect scores in predicting the composition of clay, silt, and sand.

Tom: Exactly! That high accuracy with KNN suggests that there's a really strong mathematical link between those light measurements and the actual physical makeup of the soil particles. It moves us beyond just 'this looks like clay' to actually quantifying *how much* clay is in there.

Jane: And it’s not just about classification; they also showed that you can use this system to predict the exact percentage breakdown of clay, silt, and sand with very high precision, which is super useful for anyone planning an agricultural operation.

Lu: The implication here for my work is huge because it validates using spectral data as a reliable proxy for complex physical properties like particle size distribution that have historically been difficult to measure non-destructively across large areas.

Meng: So, if we take this into the real world, it means we could potentially set up systems that screen large tracts of land rapidly, which cuts down on the time and money usually spent on detailed soil surveys.

Lalam: That’s right; for culture and science, this kind of tool can really democratize access to detailed soil characterization, making high-level scientific knowledge accessible to field workers and local governments much faster than traditional methods allow.

Tom: It’s exciting because it bridges that gap between the slow, expensive lab tests and scalable, data-driven assessment right there in the field. So now we know *how* they did it; what does this actually mean for how we use this technology in the next few years?

The paper's summary: Jane: Now that we understand how they achieved those strong results, let’s talk about what the authors suggest to make the "Reflectance Multispectral Imaging for Soil Composition Estimation and USDA Texture Classification" even better in the future.

Lu: They suggest integrating a comparative decision support module that lets you compare the results of direct classification against those derived from indirect composition mapping, which is really smart for understanding when to trust one method over the other.

Meng: That makes sense; having a system that can tell you whether to go for a fast single-step screening or a more detailed compositional analysis based on your needs is exactly what we need for practical engineering applications.

Lalam: I think the authors also mention that focusing on hardening the hardware against environmental noise and variability will be crucial because field conditions aren't always perfectly controlled like in their dark chamber tests.

Tom: Right, so they’re looking at making this tool more resilient to real-world messiness, which is a necessary step before you can deploy it widely in agriculture or geotechnical screening.

Jane: And they also touch on the importance of refining the feature extraction process itself, suggesting that perhaps exploring different ways to partition those image blocks could yield even better results for capturing subtle soil differences.

Lu: I think those suggestions are really exciting because they open up a lot of avenues for future research; thinking about how these learned feature spaces can inform new generative models in material science is where things get really imaginative.

Meng: From an engineering standpoint, if they can solidify the hardware and the software pipeline to be more adaptable, we could see this capability move from a proof-of-concept device to a standardized tool that works across different geographic regions.

Lalam: I think the most impactful vision here is how this advance can improve culture by providing accessible, high-fidelity data on soil composition, which helps local communities and researchers understand their natural resources much better.

Tom: It really does; moving from just a proof-of-concept to a reliable tool for field screening means we’re getting closer to a world where soil characterization isn't just a slow, expensive process anymore. So, if this hardware gets hardened and the software refined as suggested, what do you think is the next big thing we should look at?

The paper's improvements: Jane: So we’ve covered how this paper tackles soil texture prediction using multispectral imaging and machine learning, and now we're wrapping up our discussion on "Reflectance Multispectral Imaging for Soil Composition Estimation and USDA Texture Classification."

Tom: That's right; we saw how they built a system that lets us translate light signatures into concrete information about soil type, from clay to sand. It’s a really practical application of AI in the physical sciences.

Lu: I think the big picture here is that this work shows how we can use existing imaging hardware to solve long-standing classification challenges in agriculture and engineering, which is a fascinating path for future multimodal research.

Meng: From my perspective, this means we could see a huge reduction in the time and cost associated with detailed soil surveys because you get reliable data directly from the field. That kind of efficiency is exactly what I look for when evaluating new AI tools.

Lalam: The most impactful vision I have is how this advance can improve culture by providing accessible, high-fidelity data on soil composition, which helps local communities and researchers understand their natural resources much better than ever before.

Tom: It really does; we’re moving away from slow, manual methods toward a scalable process that’s ready for deployment right now.

Jane: And remember what we learned about the different classification strategies—direct versus indirect—which gives us flexibility depending on whether we need a quick answer or a detailed breakdown of composition.

Lu: Indeed; the potential for this technology is vast, and I think applying these same MSI principles to characterizing other materials or even monitoring soil health changes over time with high fidelity is where things get really creative.

Meng: I’ll be looking closely at how they harden this system for harsh field conditions so we can transition this from a successful experiment to a truly robust tool.

Lalam: What we’ve discussed today about "Reflectance Multispectral Imaging for Soil Composition Estimation and USDA Texture Classification" really highlights how powerful AI can be when applied to physical sciences to create tools that make a tangible difference in our daily work.

Tom: It’s been great exploring these results with you all; this research proves that combining specialized hardware with robust machine learning techniques can create tools that are both highly accurate and incredibly practical for real-world deployment.

Jane: I agree; it’s a fantastic example of how traditional knowledge, like the USDA texture triangle, can be successfully integrated with modern data science to create something useful.

Lu: The next step involves exploring how these learned feature spaces can inform new generative models in material science, which is where things get really imaginative.

Meng: I’ll be looking closely at the system design details to see how we can integrate this capability into our existing sensor platforms for rapid prototyping.

Conclusion: Tom: So we’ve covered how they tackled soil texture prediction using reflectance multispectral imaging and machine learning in "Reflectance Multispectral Imaging for Soil Composition Estimation and USDA Texture Classification."

Jane: That's right, Tom; we saw how they built a system that lets us translate light signatures into concrete information about soil type, from clay to sand. It’s a really practical application of AI in the physical sciences.

Lu: I think the big picture here is that this work shows how we can use existing imaging hardware to solve long-standing classification challenges in agriculture and engineering, which is a fascinating path for future multimodal research.

Meng: From my perspective, this means we could see a huge reduction in the time and cost associated with detailed soil surveys because you get reliable data directly from the field. That kind of efficiency is exactly what I look for when evaluating new AI tools.

Lalam: The most impactful vision I have is how this advance can improve culture by providing accessible, high-fidelity data on soil composition, which helps local communities and researchers understand their natural resources much better than ever before.

Tom: It really does; we’re moving away from slow, manual methods toward a scalable process that’s ready for deployment right now.

Jane: And remember what we learned about the different classification strategies—direct versus indirect—which gives us flexibility depending on whether we need a quick answer or a detailed breakdown of composition.

Lu: Indeed; the potential for this technology is vast, and I think applying these same MSI principles to characterizing other materials or even monitoring soil health changes over time with high fidelity is where things get really creative.

Meng: I'll be looking closely at how they harden this system for harsh field conditions so we can transition this from a successful experiment to a truly robust tool.

Lalam: What we’ve discussed today about "Reflectance Multispectral Imaging for Soil Composition Estimation and USDA Texture Classification" really highlights how powerful AI can be when applied to physical sciences to create tools that make a tangible difference in our daily work.

Tom: It’s been great exploring these results with you all; this research proves that combining specialized hardware with robust machine learning techniques can create tools that are both highly accurate and incredibly practical for real-world deployment.

Jane: I agree; it’s a fantastic example of how traditional knowledge, like the USDA texture triangle, can be successfully integrated with modern data science to create something useful.

Lu: The next step involves exploring how these learned feature spaces can inform new generative models in material science, which is where things get really imaginative.

Meng: I’ll be looking closely at the system design details to see how we can integrate this capability into our existing sensor platforms for rapid prototyping.

G.A.S.L. RANASINGHE, J.A.S.T. JAYAKODY, M.C.L., DE SILVA, G., THILAKARATHNE, G., M., , , , ,

Department of Electrical and Electronic Engineering, University of Peradeniya · Multidisciplinary AI Research Center, University of Peradeniya · Department of Electrical and Information Engineering, University of Ruhuna · Department of Civil Engineering, University of Peradeniya

cs.CV, eess.SP

Submitted: 2026-02-26

Updated: 2026-02-26

Comments: Under Review at IEEE Access. 17 pages, 15 figures

Journal ref: IEEE Access, vol. 14, pp. 128744-128763, 2026

DOI: 10.1109/ACCESS.2026.3721276

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 84/100

The gist: This manuscript proposes a robust and field deployable multispectral imaging (MSI) system combined with machine learning to accurately predict soil composition and the United States Department of

Key concepts

Multispectral Imaging (MSI)
This involves capturing light reflectance across thirteen specific wavelengths (365 nm to 940 nm) using a specialized camera and LEDs. This allows the system to gather rich spectral data about the soil's composition, which is then processed by machine learning models.
Feature Extraction
The raw spectral data is transformed into meaningful features. The researchers divided the image into a grid and calculated the average intensity for each wavelength band within those blocks. This process creates a numerical matrix representing the soil's spectral signature for machine learning to analyze.
Soil Texture Classes
These are twelve distinct categories used by the USDA to classify soil based on particle size, such as clay, silt, and sand. The study aimed to use MSI data to accurately predict which of these specific texture classes a soil sample belongs to.
Machine Learning Strategies
The study tested three ways to use the extracted features: directly classifying textures, regressing the percentages of clay/silt/sand, or indirectly mapping those compositions onto texture classes using USDA rules. The results showed that direct classification was the most accurate approach.

Terminology

Summary

This manuscript proposes a robust and field deployable multispectral imaging (MSI) system combined with machine learning to accurately predict soil composition and the United States Department of Agriculture (USDA) texture classes. This research is significant because it offers a cost-effective, non-destructive method for soil characterization, overcoming the limitations of slow laboratory particle size tests while providing data suitable for geotechnical screening and precision agriculture.

The Proposed System and Data Acquisition

The system utilizes a cost effective in-house MSI device operating from 365 nm to 940 nm to capture thirteen spectral bands. This setup uses a FLIR BFS-U3-13Y3M CMOS monochrome machine vision camera and an array of narrowband LEDs at thirteen central wavelengths. The acquisition is performed within a custom dark chamber with nonreflective matte black material to minimize interference from ambient light and external reflections, ensuring repeatable measurements.

Soil Sample Preparation and Ground Truth

The study utilized three representative source soils—clay rich soil, silt rich soil, and sand rich soil—as endmembers. These were mixed in controlled mass ratios to cover all twelve USDA soil texture classes. A total of 524 specimens were prepared for the dataset (training/testing), involving 22 unique mixture ratios with 20 replicates each for training/testing, and 7 additional mixtures for external validation. Ground truth particle size composition was obtained through laboratory tests using sieve analysis and hydrometer analysis, following ASTM D7928-21e01 standards.

Image Preprocessing and Feature Extraction

Each multispectral sample undergoes a standardized preprocessing pipeline:

  1. Dark current correction using an absolute difference operation: Y (λ) = X(λ) − D.

  2. Region of Interest (ROI) selection and cropping to a fixed 100 × 100 pixel area.

  3. Contrast normalization applied via a bounded non-linear intensity mapping using the hyperbolic tangent function: Yˆomega(λ, p) = (µλ − σλ)+2σλ·tanh(κ (Yomega(λ, p) − µλ)) + 1/2.

Spectral features are then extracted by partitioning the ROI into a regular grid of 10×10 non-overlapping blocks. The mean intensity of each block for every wavelength band is computed to form a per sample feature matrix, resulting in a feature set of X = [x1, x2,..., x13] ∈ R100×13.

Machine Learning Framework and Classification Strategies

The framework evaluates three complementary strategies:

(i) Direct classification of the twelve USDA texture classes from multispectral features.

(ii) Regression to estimate clay, silt, and sand percentages from the same features.

(iii) Indirect classification by mapping the regressed compositions to texture classes via the USDA soil texture triangle decision rules.

Dimensionality reduction is performed using Linear Discriminant Analysis (LDA), which projects the 13-dimensional feature space into a lower-dimensional subspace (K=5 components) to enhance class separability. The final models tested include K-Nearest Neighbors (KNN), Random Forest (RF), Decision Trees (DT), CatBoost (CB), and XGBoost.

Key Performance Results

The evaluation demonstrated strong performance across all strategies:

(i) Direct Classification:

KNN achieved the best overall performance with a mean accuracy of 0.9955, while RF and XGB delivered comparable results, reaching accuracies around 0.9942–0.9951.

(ii) Composition Regression:

KNN achieved the best predictive accuracy across all components: Clay (R2 = 0.9993, RMSE = 0.5300), Silt (R2 = 0.9988, RMSE = 0.9159), and Sand (R2 = 0.9982, RMSE = 1.1747).

(iii) Indirect Classification:

KNN achieved the highest accuracy of mean accuracy of 0.9698, comparable to the direct method, although a slight drop was noted due to the two-step process and linear boundaries of the USDA soil texture triangle.

The study concludes that this MSI feature space is highly discriminative for texture assessment, enabling accurate prediction and supporting a practical bridge between classical soil texture knowledge and scalable, data-driven soil assessment. The direct approach consistently outperformed the indirect pipeline by 0.0257 in absolute accuracy due to the discontinuous nature of the triangle mapping. Additionally, KNN was found to be particularly effective for composition estimation because the relationship between multispectral signatures and soil composition is well captured by local structure in the feature space. The results confirm that MSI provides a viable, highly scalable tool for rapid, low cost field screening.

Improvements for AI systems

Here are the specific improvements for AI systems based on this research, and what those improved systems can achieve:

) Improved AI System 1: Cost-Effective, Field-Deployable Soil Texture Characterization System (MSI + End-to-End Learning Pipeline)

This system integrates the proposed hardware and software pipeline.

  1. A custom, low-cost MSI device operating from 365 nm to 940 nm captures 13 spectral bands using narrowband LEDs, optimized for soil reflectance.

  2. An end-to-end machine learning framework (combining LDA dimensionality reduction and various classifiers like KNN, RF, XGB) is trained on a curated dataset of laboratory-prepared mixtures (clay/silt/sand combinations).

This improved system can:

  • Perform non-destructive soil texture characterization in the field with low hardware costs.

  • Estimate soil composition (clay, silt, sand percentages) with high accuracy (e.g., R2 up to 0.9993 for clay).

  • Directly classify soils into one of the twelve USDA textural classes (achieving over 99% accuracy).

) Improved AI System 2: Predictive Soil Composition Regression Model

This system focuses on the regression pipeline, specifically leveraging the most accurate regressor identified.

  1. The system uses a trained regression model (e.g., KNN, achieving R2 > 0.986 for all components) to map the captured MSI spectral features directly to soil composition percentages (clay, silt, sand).

This improved system can:

  • Provide precise quantitative estimates of clay, silt, and sand fractions from raw spectral data.

  • Offer highly reliable input parameters for subsequent agronomic models.

) Improved AI System 3: Indirect Soil Texture Classification via Composition Mapping Module

This system uses the predictive power of the composition model to infer texture classes without direct classification training on texture labels.

  1. The system takes the predicted (clay, silt, sand) triplet from System 2 and applies a rule-based mapping against the USDA soil texture triangle decision rules.

This improved system can:

  • Classify soil textures by leveraging composition estimates rather than relying solely on direct spectral pattern matching.

  • Provide a mechanism for interpreting the physical meaning of spectral features in terms of established soil science knowledge (the texture triangle).

) Improved AI System 4: Comparative Decision Support Module (Direct vs. Indirect Strategy Selector)

This meta-system compares the outputs of System 1's Direct Classification and System 3's Indirect Classification pipelines.

  1. The module calculates a performance metric (e.g., mean accuracy) for both strategies on new, unseen data and provides a recommendation based on operational goals (speed vs. interpretability).

This improved system can:

  • Determine the optimal classification strategy for a given application: selecting the fast, single-step Direct Classification when rapid screening is needed, or selecting the more interpretable Indirect Classification when detailed compositional estimates are required for management decisions.

Related papers