GreenGuardian: Computer Vision on Embedded Edge Devices

The Agricultural Challenge

The agricultural industry is essential for feeding the world, yet weed management remains one of its most expensive and ecologically damaging challenges. In 2021 alone, global agriculture consumed over 1.7 million metric tons of chemical herbicides.

Broadcast spraying blankets entire fields with toxic chemicals to kill weeds that only occupy a fraction of the soil surface. This leads to:

• Soil microbiome degradation and long-term land infertility.
• Herbicide runoff into local groundwater tables and municipal drinking supplies.
• Herbicide-resistant superweeds, forcing farmers to use progressively harsher chemical cocktails.

Small-scale farmers are hit hardest: industrial mechanical weeding machines cost hundreds of thousands of dollars, leaving them with no viable alternative to chemical spraying.

GreenGuardian was built to solve this: an affordable, autonomous agricultural rover that uses computer vision to identify invasive weeds in real time and activate a targeted solenoid spray nozzle directly above the weed — cutting herbicide volume by ~90%.

---

Hardware Prototype Iterations

Building a rover capable of traversing uneven farm furrows, tall grass, and muddy soil required three major mechanical iterations:

Iteration 1: Salvaged toy DC motors + single servo steering
  -> Problem: Severe lack of torque; bogged down immediately in grass.

Iteration 2: 3D-printed rack-and-pinion front steering
  -> Problem: High mechanical friction on rough terrain; steering servo stripped under load.

Iteration 3 (Final): Dual independent high-torque rear motors + skid steering + 3D printed chassis
  -> Result: 360-degree zero-radius turning, robust traversal across inclines and bumps.

---

Computer Vision & Edge AI Pipeline

[Camera Stream] 
       |
       v
[Adaptive Histogram Equalization (CLAHE)]  <-- Cancels harsh direct sun / shadows
       |
       v
[HSV Plant Segmentation Mask]              <-- Isolates green vegetation from soil
       |
       v
[TFLite 8-bit Quantized YOLOv5 Model]      <-- Classifies Crop vs. Invasive Weed (<150ms)
       |
       v
[Actuator Coordinate Mapper] 
       |
       v
[Targeted Solenoid Pulse (0.15s)]          <-- 90% chemical reduction

1. Eliminating Sunlight Distortion

Outdoors, illumination varies wildly between direct harsh sunlight (which blows out green channels) and crop canopy shadows (which turn plants nearly black).

To solve this, I designed a two-stage preprocessing pipeline:

1. Contrast Limited Adaptive Histogram Equalization (CLAHE) applied across the luminance channel.
2. HSV Color-Space Thresholding to create a robust invariant vegetation mask before neural network inference.

2. PyTorch to 8-Bit TFLite Quantization

Running a standard FP32 YOLOv5 model on a Raspberry Pi 4 resulted in ~1.2 seconds of latency per frame — far too slow for real-time robotic navigation.

By quantizing the trained PyTorch weights into 8-bit integer TFLite models, inference dropped to <150ms per frame on the Pi's CPU with negligible drop in classification precision.

import RPi.GPIO as GPIO
import time

SPRAY_PIN = 18
GPIO.setmode(GPIO.BCM)
GPIO.setup(SPRAY_PIN, GPIO.OUT)

def trigger_targeted_spray(x, y, confidence):
    if confidence > 0.85:
        # Weed localized within actuator blast zone
        GPIO.output(SPRAY_PIN, GPIO.HIGH)
        time.sleep(0.15)
        GPIO.output(SPRAY_PIN, GPIO.LOW)

---

National Recognition & Results

GreenGuardian was evaluated on test plots with invasive dandelions, thistle, and broadleaf weeds:

• 94% weed detection precision under varying outdoor lighting.
• ~90% reduction in herbicide volume compared to standard broadcast treatment.
• Awarded Bronze Medal at the 2024 Canada-Wide Science Fair (CWSF) in Ottawa.

Check out the full project details on the CWSF ProjectBoard.