AS7263 NIR · ESP32 · Random Forest ML

SpectrOver

A near-infrared spectroscopy system that identifies material composition — Paper, Wood, Plastic, Metal — in real time.

610nm

680nm

730nm

760nm

810nm

860nm

Live spectral fingerprint — Paper sample @ 33°C

THE TEAM

Built by students,
driven by curiosity.

Three students constructed this end-to-end spectroscopy pipeline from scratch — designing the cardboard enclosure, wiring the ESP32 sensor, collecting training data, and writing the ML classifier entirely themselves.

Material classes

Engineered features

NIR channels

Architecture

How It Works

🔦

NIR Illumination

An LED at 25 mA illuminates the target material. The AS7263 sensor captures reflected light across 6 near-infrared channels (610–860 nm), producing a unique spectral fingerprint for each material type.

⚙️

Feature Engineering

Raw channel values are expanded into 21 features: normalized ratios, inter-channel gradients (slopes), and statistical descriptors — capturing the shape of each reflection curve, not just its amplitude.

🧠

Random Forest Inference

A 150-tree Random Forest classifier, trained on hundreds of labelled samples with oversampling for class balance, identifies the material in real time over a live serial stream from the ESP32.

AS7263 NIR sensor channel readings for a sample

Live Data

Raw Sensor Readings

The AS7263 outputs calibrated intensity values across six NIR bands simultaneously. Channel R at 610 nm dominates for most materials, while the ratio between channels encodes the material identity. Gain: 64×, Integration time: 168 ms.

R 610 nm S 680 nm T 730 nm U 760 nm V 810 nm W 860 nm

Computer Vision

Colour Spectrum Analysis

A region-of-interest (ROI) is extracted from the camera feed and analysed for RGB channel gradients across pixel positions. This visual fingerprint complements the NIR spectral data, providing an additional discriminative signal for material classification.

Source

Feature Engineering Pipeline

ml_train.py — engineer_features()

def engineer_features(X_raw):
    """
    Transforms raw spectroscopy channels into relational features
    to capture the 'shape' of the material reflection curve.
    """
    X_eng = pd.DataFrame(index=X_raw.index)
    raw_channels = ['R', 'S', 'T', 'U', 'V', 'W']

    # 1. Retain the original raw values
    for col in raw_channels:
        X_eng[col] = X_raw[col]

    # 2. Total Spectral Footprint Intensity
    X_eng['Total_Intensity'] = X_raw[raw_channels].sum(axis=1).replace(0, 1)

    # 3. Normalized Curves (Ratios relative to total brightness)
    for col in raw_channels:
        X_eng[f'{col}_norm'] = X_raw[col] / X_eng['Total_Intensity']

    # 4. Spectral Gradients (Slopes between consecutive channels)
    X_eng['R_S_diff'] = X_raw['R'] - X_raw['S']
    X_eng['S_T_diff'] = X_raw['S'] - X_raw['T']
    X_eng['T_U_diff'] = X_raw['T'] - X_raw['U']
    X_eng['U_V_diff'] = X_raw['U'] - X_raw['V']
    X_eng['V_W_diff'] = X_raw['V'] - X_raw['W']

    # 5. Row-wise Statistical Profiles
    X_eng['Spectral_Mean'] = X_raw[raw_channels].mean(axis=1)
    X_eng['Spectral_Std']  = X_raw[raw_channels].std(axis=1)
    X_eng['Spectral_Max']  = X_raw[raw_channels].max(axis=1)

    return X_eng