Installation and Setup

The system is designed to run in a local Python environment.

1. Clone the Repository

Download the source code to your local machine: go to your vscode terminal and type Bash

git clone https://github.com/pojeesahil/DS.git
cd DS

2. Environment Configuration

Install the necessary dependencies using the provided requirements file: Bash

pip install -r requirements.txt

install jupyter notebook extension in vscode preferably

Requirements:

earthengine-api: Google Earth Engine interface.

geemap: Geospatial analysis and visualization.

torch: PyTorch framework for neural network training.

numpy: Numerical data processing.

matplotlib: Results visualization and mapping.

3. Google Earth Engine Authentication

The project requires access to Google’s Earth Engine API. Upon the first execution of the script, a browser window will open to authorize the application.

Log in with a Google account associated with a registered Earth Engine project.

Select the appropriate Cloud Project ID when prompted.

Run the analysis using either: 1)manually running predict.ipynb in vscode 2)Running predict.ipynb on google collab in broswer

Soil Erosion Detection using RUSLE, Time-Series Satellite Data, and Deep Learning

This repository provides a quantitative approach to identifying soil erosion risk by integrating the Revised Universal Soil Loss Equation (RUSLE) with multi-year time-series satellite imagery and deep learning. By leveraging Google Earth Engine for environmental data and PyTorch for spatial analysis, this system evaluates terrain to classify genuine landscape degradation risks.

Overview

The project operates on a hybrid "Theory plus Reality" model. It calculates a scientific baseline for soil loss based on environmental factors (theoretical risk) and cross-references this with a five-year historical analysis of actual vegetation loss (observed reality). This dual-verification data trains a U-Net convolutional neural network to identify high-risk areas from multispectral satellite data and topographic slope, filtering out seasonal noise.

Technical Methodology

The system quantifies soil loss ($A$) using the standard RUSLE formula:

$$A = R * K * LS * C * P$$

• R (Rainfall Erosivity): Derived from CHIRPS daily precipitation data to measure the kinetic energy of rainfall.
• K (Soil Erodibility): Calculated from OpenLandMap soil texture classes, representing soil particle detachment susceptibility.
• LS (Slope Length and Steepness): Generated from the USGS SRTM Digital Elevation Model to account for topographic effects on runoff velocity.
• C (Cover Management): Calculated using NDVI from Sentinel-2 imagery to evaluate surface protection by vegetation.
• P (Support Practices): Evaluated as a constant, assuming natural land management.

Split-Logic Processing

To ensure accuracy across diverse landscapes, the model employs specialized logic based on the ESA WorldCover dataset:

• Wildlands and Forests: Classification requires both a severe mathematical RUSLE score and a significant historical drop in NDVI (10%+).
• Agricultural Lands: To avoid false alarms caused by standard crop harvesting cycles, farmlands are evaluated strictly on physical topography and rainfall thresholds.
• Masking: Urban infrastructure, permanent water bodies, and snow are excluded to prevent statistical bias.

Output Interpretation

The analysis generates a three-pane visualization:

1. Satellite View: A true-color RGB composite of the selected region.
2. Theoretical RUSLE Baseline: The raw mathematical calculation highlighting all vulnerable slopes.
3. Hybrid AI Inference: The finalized neural network prediction pinpointing actual, confirmed degradation.

Risk Classification

• Green (Stable): Minimal calculated soil loss or stable vegetation cover.
• Yellow (Vulnerable): High-tension mathematical buffer zones acting as a warning tier.
• Red (Critical): High-risk zones requiring immediate intervention, exhibiting both topographical vulnerability and proven historical degradation.

Level 3 Mitigation Advisory

The system includes a targeted intervention feature. When critical (Red) zones are detected, the model:

• Identifies the largest contiguous erosion cluster using center-of-mass calculations.
• Drops a dynamic GPS marker on the interactive map.
• Provides engineering recommendations (e.g., Gabion walls for steep slopes vs. riparian buffers for low slopes).