Spectra-to-Image Deep Learning Project

📌 Overview

This project, developed under Sci-Ware, explores spectral data transformation into 2D image representations and applies deep learning models (CNNs) for regression tasks such as predicting Moisture and other chemical properties. Additionally, a PLS (Partial Least Squares) baseline model is implemented for comparison with traditional chemometrics.

---

🎯 Objectives

• Convert 1D spectral data into structured 2D images using different reshaping techniques:

  • Raw Reshape (Row-major order)
  • Column-major Reshape
  • Snake Pattern Reshape

• Apply preprocessing methods:

  • Transmission → Absorbance
  • Derivative transformation (Savitzky-Golay, 2nd order)
  • Mean centering / Standard Normalization (Z-score)
  • Resampling spectra to fixed dimensions

• Build deep learning pipelines:

  • 2D CNN model (inspired by VGG-like architecture)
  • Compare with PLS baseline using SIMPLS algorithm

• Evaluate performance using:

  • RMSEC, RMSECV, RMSEP
  • R² (Calibration, Cross-validation, Prediction)
  • Bias analysis

---

🧪 Dataset

• X-block (Spectra):

  • Samples: 432
  • Variables: 257 spectral points
  • Preprocessing: Transmission → Absorbance, 2nd derivative, mean centering

• Y-block (Targets):

  • Example target: Moisture (%)
  • Format: CSV file with one column per property

⚠️ Dataset and experimental data are proprietary and belong to Sci-Ware.

---

🔄 Data Transformation

Spectral data is resampled and reshaped into 2D images:

• Snake Pattern Example:

def to_snake_pattern(vector, img_size):
    img = np.zeros((img_size, img_size))
    for i in range(img_size):
        row = vector[i * img_size:(i + 1) * img_size]
        if i % 2 == 1:
            row = row[::-1]
        img[i, :] = row
    return img

• Final image size used: 65 × 65 pixels

---

🧠 Models

🔹 1. PLS Regression (Baseline)

• Algorithm: SIMPLS

• Latent Variables: 8

• Cross-validation: Venetian blinds (5 splits)

• Example results for Moisture:

  • RMSEC = 2.1651
  • RMSECV = 2.2557
  • RMSEP = 1845.59
  • R² Cal = 0.8876
  • R² CV = 0.8780
  • R² Pred = 0.9050

---

🔹 2. CNN Model (Deep Learning)

Input: 1 × 65 × 65 image

Architecture:

Input → Conv2D(64,3×3) → MaxPool(2×2)  
      → Conv2D(128,3×3) → MaxPool(2×2)  
      → Conv2D(256,3×3) → Conv2D(256,3×3) → MaxPool(2×2)  
      → Conv2D(512,3×3) → Conv2D(512,3×3) → MaxPool(2×2)  
      → Conv2D(512,3×3) → Conv2D(512,3×3) → MaxPool(2×2)  
      → Flatten → Dense(128) → Dense(Output)

Implemented in PyTorch.

---

⚖️ Evaluation

• Comparison between PLS (classical regression) vs CNN (deep learning).
• CNN showed higher prediction accuracy on transformed spectral images compared to PLS in cross-validation.
• Remaining work: Extend CNN evaluation on all transformation methods (Raw, Column, Snake).

---

🛠️ Technologies

• Python 3.10

• Libraries:

  • PyTorch (deep learning)
  • Scikit-learn (PLS, metrics)
  • SciPy (resampling)
  • NumPy, Pandas (data handling)
  • Matplotlib/Seaborn (visualization)

---

📊 Example Visualizations

• Original spectra
• Resampled spectra
• 2D image (snake pattern)

(Add plots here if available)

---

🚀 Future Work

• Compare image reshaping techniques (Row-major vs Column-major vs Snake).
• Test CNN on larger datasets.
• Experiment with transfer learning from pretrained vision models.
• Optimize hyperparameters (batch size, learning rate, dropout).

---

👨‍💻 Author & Ownership

• Developed by Rowaina Reda and Salma Bassem
• 📅 Started: Sept 2025
• 🏢 Project under Sci-Ware

---