SoilSage: AI-Driven Precision Soil Sustainability

Overview

This is a cyber-physical system designed to combat soil acidification and nutrient runoff—major environmental concerns in modern chemistry. By integrating IoT sensors with a Random Forest Machine Learning model, the system identifies the soil's chemical state and provides precise neutralization strategies through recommendations of crops and fertilizers.

Course: ECS2001 - Project Ideation, Design and Prototyping
Created by: Rahul, Abhiram, Nikhil, Sanjeev, Saketh and Shyam

Table of Contents

• Overview
• Features
• Project Structure
• Setup
• Running the Application
  • Browser Dashboard
  • Command Line
• Flags and Options
• ML Algorithm & Training

---

Features

• Hardware Sensing: Real-time pH, NPK, Soil Temperature (DS18B20), and Air Temperature/Humidity (DHT11) monitoring via ESP32.
• Cloud Rainfall Input: 30-day rainfall is fetched online by location (Open-Meteo) for model input.
• AI Inference: Random Forest Classifier trained on agricultural datasets.
• Smart Recommendations: Crop prediction + fertilizer suggestion with pH-aware chemistry advice.

Project Structure

• app/app.py: Browser dashboard (live readings + crop + fertilizer recommendation).
• scripts/: Contains utility scripts.
  • ser_script.py: Local Python bridge between ESP32 and ML model.
  • sim_script.py: Simulation script for software-only demonstration.
• random_forest/: Contains all machine learning related files.
  • dataset/: Contains the dataset used for training.
  • model/: Contains the serialized Random Forest model.
  • training/: Contains the training Jupyter notebook.
• docs/setup.md: Required hardware components and ESP32 wiring map.
• esp32/: Directory containing ESP32 firmware files for data acquisition.

Setup

Before proceeding, set up the hardware per docs/setup.md and flash esp32/esp32_full/esp32_full.ino via ESP32 IDE.

1. Connect the ESP32 to your computer (for live/serial mode). This is commonly /dev/ttyACM0 or /dev/ttyUSB0 for Linux/macOS, or COM3 for Windows.
2. Install dependencies:

Linux / macOS:

python3 -m pip install pandas joblib pyserial scikit-learn flask

Windows:

py -m pip install pandas joblib pyserial scikit-learn flask

Running the Application

🌐 Browser Dashboard

Start the local web dashboard:

# Simulation mode (default)
python3 app/app.py

# Wi-Fi mode (listen for ESP32)
python3 app/app.py --esp32

# Serial mode (connect to ESP32 via USB)
python3 app/app.py --serial

Open http://127.0.0.1:5000 in your browser.

• Set your coordinates using Use Browser Location or manual latitude/longitude.
• The app fetches rolling 30-day rainfall (mm) online and feeds it to the model.
• If location/weather fetch is unavailable, the app falls back to cached or default rainfall and shows the source status in the UI.

Command Line

You can run the core logic purely from the terminal without the web UI:

Linux / macOS:

# Simulation mode
python3 scripts/sim_script.py

# Serial mode connected to ESP32
python3 scripts/ser_script.py --port=/dev/ttyACM0

Windows:

# Simulation mode
py scripts/sim_script.py

# Serial mode connected to ESP32
py scripts/ser_script.py --port=COM3

Flags and Options

App flags (app/app.py)

• app/app.py → starts in SIM mode
• --sim → explicitly starts in SIM mode
• --serial <PORT> → starts in SERIAL mode
• --serial → starts in SERIAL mode and auto-detects the port if /dev/ttyACM0 is unavailable
• --esp32 → starts the app in Wi-Fi Mode listening on 0.0.0.0 for POST requests from the ESP32.
• --lock-mode → hides the mode selector and prevents switching modes during runtime
• --no-check → skips strict pH input range validation (chemical status still shown)

Note: The mode selector can be hidden by pressing 'm' in normal mode.

Script Flags (scripts/sim_script.py, scripts/ser_script.py)

These scripts support optional flags to provide rainfall data:

• --location <LAT> <LON>: Fetches 30-day rainfall data from Open-Meteo based on the provided latitude and longitude.
  • If the API call fails, the script falls back to --rainfall_data, or a default value (100.0mm).
• --rainfall_data <VALUE>: Directly specifies the 30-day rainfall value in millimeters.

Examples:

# Simulation with location-based rainfall
python3 scripts/sim_script.py --location 34.0522 -118.2437

# Serial with direct rainfall data
python3 scripts/ser_script.py --port=/dev/ttyACM0 --rainfall_data 150.0

ML Algorithm & Training

Model Details

• Algorithm: RandomForestClassifier
• Input features (strict order): Nitrogen, phosphorus, potassium, temperature, humidity, ph, rainfall
• Target label: label (crop name)
• Train/test split: 80/20 using train_test_split(..., test_size=0.2, random_state=42)
• Model params: n_estimators=100, random_state=42
• Saved model artifact: random_forest/model/soil_model.pkl (via joblib.dump)

Training Environment

• Trained in Jupyter Notebook.
• Training code is located in random_forest/training/train.ipynb.
• Dataset schema follows random_forest/dataset/Crop_recommendation.csv.

Using the trained model in this project

1. Place the generated soil_model.pkl in the random_forest/model/ directory.
2. Ensure runtime feature order exactly matches training feature order.
3. Run scripts/ser_script.py, scripts/sim_script.py, or app/app.py to load the model and predict crops.