SolarQuant: A Self-Learning Hybrid AI System

Overview

SolarQuant is a next-generation solar potential analysis pipeline developed for the EcoInnovators Ideathon 2026. It addresses the limitations of traditional single-model computer vision systems by employing a Hybrid Intelligence Architecture. This system synthesizes the speed of YOLOv8 segmentation, the semantic understanding of Vision Transformers (ViT), and the geometric precision of classical Computer Vision to deliver highly accurate rooftop solar assessments.

Team: Streetwise V78349

---

Directory Structure

This project adheres to the strict delivery compliance standards set forth in the Challenge Overview.

• Pipeline code/: Core application logic including main.py (Entry Point) and solar_model.py (Inference Logic).
• Environment details/: Dependency manifests (requirements.txt, environment.yml, python_version.txt).
• Trained model file/: Contains the fine-tuned model weights (.pt format).
• Model card/: Comprehensive documentation of model dataset, logic, limitations, and bias.
• Prediction files/: Generated JSON output containing analysis results.
• Artefacts/: Visual verification data including segmentation overlays.
• Model Training Logs/: Performance metrics and training history.
• README.md: Technical documentation and execution guide.

---

Quick Start Guide

1. Installation

Install the required dependencies using the provided requirements file:

pip install -r "Environment details/requirements.txt"

⚠️ Please configure HuggingFace and Google Maps API in the config.py File before running

2. Execution

Navigate to the pipeline directory and execute the main script:

cd "Pipeline code"
python app.py
Access the App on Local server @ http://127.0.0.1:5003

• Input: The system processes Pipeline code/input_data.xlsx.
• Output: Results are generated in Prediction files/results.json. Visual overlays are saved to Artefacts/overlays.

3. Continuous Learning

SolarQuant features an automated self-evolution mechanism. When the Hybrid Fallback system identifies panels missed by the primary model, these examples are automatically curated for retraining. To trigger the model update process:

python train_yolo.py

Subsequent executions of main.py will utilize the evolved model, progressively reducing false negatives over time.

---

Methodology

Hybrid Architecture

1. Semantic Guardrail: Utilizes google/vit-base-patch16-224 (Hugging Face) to classify image content and detect solar presence contextually.
2. Instance Segmentation: Deploys finloop/yolov8s-seg-solar-panels for precise boundary detection and area calculation.
3. Geometric Fallback: Implements OpenCV-based edge detection and adaptive thresholding to identify missed panels in high-confidence scenarios.

Buffer Analysis Logic

The system implements a hierarchical buffer analysis protocol:

1. Primary Zone (1200 sq.ft): Prioritized scan within a ~6m radius.
2. Secondary Zone (2400 sq.ft): Extended scan within a ~8.5m radius if the primary scan yields negative results.
3. Strict Compliance: Detections falling outside these zones are filtered to maintain specific site relevance.

---

License and Attribution

• License: MIT License
• Data Source: Google Maps Static API
• Base Models: Ultralytics YOLOv8 (AGPL-3.0), Google ViT (Apache 2.0).