Reflectance Multispectral Imaging for Soil Composition Estimation and USDA Texture Classification

Soil texture is a foundational attribute that governs water availability and erosion in agriculture, as well as load bearing capacity, deformation response, and shrink-swell risk in geotechnical engineering. Yet texture is still typically determined by slow and labour intensive laboratory particle size tests, while many sensing alternatives are either costly or too coarse to support routine field scale deployment. This paper proposes a robust and field deployable multispectral imaging (MSI) system and machine learning framework for predicting soil composition and the United States Department of Agriculture (USDA) texture classes. The proposed system uses a cost effective in-house MSI device operating from 365 nm to 940 nm to capture thirteen spectral bands, which effectively capture the spectral properties of soil texture. Regression models use the captured spectral properties to estimate clay, silt, and sand percentages, while a direct classifier predicts one of the twelve USDA textural classes. Indirect classification is obtained by mapping the regressed compositions to texture classes via the USDA soil texture triangle. The framework is evaluated on mixture data by mixing clay, silt, and sand in varying proportions, using the USDA classification triangle as a basis. Experimental results show that the proposed approach achieves a coefficient of determination R^2 up to 0.99 for composition prediction and over 99% accuracy for texture classification. These findings indicate that MSI combined with data-driven modeling can provide accurate, non-destructive, and field deployable soil texture characterization suitable for geotechnical screening and precision agriculture.

💡 Research Summary

The manuscript presents a low‑cost, field‑deployable multispectral imaging (MSI) system combined with machine‑learning models to estimate soil texture composition (percentages of clay, silt, and sand) and to classify soils into the twelve USDA texture classes. Recognizing that traditional particle‑size analysis is labor‑intensive and that existing sensing alternatives either lack sufficient spectral resolution (RGB imaging) or are prohibitively expensive (hyperspectral spectroscopy), the authors develop a custom MSI platform that captures reflectance at thirteen discrete wavelengths ranging from the near‑ultraviolet (365 nm) to the near‑infrared (940 nm). The selected bands (365, 405, 473, 530, 575, 621, 660, 735, 770, 830, 850, 890, 940 nm) were chosen to span the spectral regions most sensitive to mineralogical and particle‑size differences in soils, while keeping hardware simple and affordable.

Soil samples were sourced from three distinct locations in Sri Lanka to serve as end‑members: a clay‑rich soil (Menikhinna), a silt‑rich soil (Gelioya), and a sand‑rich soil (Chavakachcheri). After oven‑drying, sieving, and confirming their true particle‑size distributions via standard sieve and hydrometer methods, the authors created 22 mixture ratios covering the entire USDA texture triangle, with 20 replicates per ratio (440 specimens). An additional external‑validation set of 7 intermediate ratios, each with 12 replicates (84 specimens), was also prepared, yielding a total of 524 samples. Each specimen was placed in a standardized aluminum container and imaged inside a dark chamber to eliminate ambient light. Illumination and image acquisition were synchronized using an Arduino Due controller, and images were captured with a FLIR BFS‑U3‑13Y3M monochrome camera (10‑bit, 1280 × 1024). For each wavelength, the mean reflectance over the sample area was extracted, forming a 13‑dimensional feature vector.

Three modeling pathways were explored:

1. Direct Classification – Multiclass classifiers (Support Vector Machine, Random Forest, XGBoost) were trained directly on the 13‑band reflectance vectors to predict one of the twelve USDA texture classes.
2. Regression of Composition – Separate regression models (Linear Regression, Support Vector Regression, Gradient Boosting, XGBoost) were trained to predict the continuous percentages of clay, silt, and sand from the same spectral features.
3. Indirect Classification – The predicted composition from the regression models was mapped onto the USDA texture triangle using the standard decision rules, thereby yielding a class label.

Performance metrics demonstrate that the regression models achieve coefficients of determination (R²) between 0.98 and 0.99 for all three components, indicating near‑perfect reconstruction of the ground‑truth composition. The best regression model (XGBoost) also exhibits low mean absolute errors (< 1 % for each fraction). Direct classification with XGBoost reaches an overall accuracy of 99.2 %, while the indirect classification (composition → triangle) attains slightly higher accuracy (≈ 99.5 %). Confusion matrices reveal that misclassifications are confined to neighboring texture classes, reflecting the inherent continuity of the underlying particle‑size space.

The authors discuss the trade‑offs between the direct and indirect approaches. Direct classification is computationally efficient and requires a single model, but it is limited to the predefined set of classes and must be retrained if new classes are introduced. The regression‑based indirect route provides continuous composition estimates, enabling finer discrimination, interpolation, and the possibility to define custom texture categories beyond the USDA scheme. Moreover, the composition output can be integrated into hydrological or geotechnical models that require quantitative fractions rather than categorical labels.

Limitations of the study include the controlled laboratory environment (dry samples, constant illumination) and the focus on three pure end‑members. The impact of soil moisture, organic matter content, surface roughness, and field lighting variability on the MSI signatures remains to be quantified. Future work is suggested to incorporate moisture correction algorithms, expand the dataset to include a broader range of natural soils, and evaluate the system on a mobile platform for in‑situ field measurements.

In conclusion, the paper demonstrates that a modestly priced, thirteen‑band MSI system, when paired with robust machine‑learning pipelines, can deliver highly accurate, non‑destructive estimates of soil texture composition and USDA class labels. This capability holds promise for precision agriculture (optimizing irrigation and fertilization), geotechnical engineering (assessing shrink‑swell potential and bearing capacity), and environmental monitoring (soil carbon and moisture dynamics), offering a practical alternative to traditional laboratory analyses and expensive hyperspectral equipment.