TLDR: This research paper details an AI-powered method for accurately assessing urban green canopy coverage using computer vision techniques on high-resolution aerial drone imagery. Focusing on Jamshedpur, an industrial city, the study outlines a robust data processing pipeline for large TIFF files and compares image segmentation and object detection approaches. The segmentation model, leveraging YOLOv5, was found to be more effective due to its pixel-level accuracy and better handling of complex canopy shapes. Implemented on a cloud platform, this methodology provides crucial data for sustainable urban planning, environmental monitoring, and climate change mitigation.
Understanding and managing urban green spaces is vital for healthy cities, especially in the face of climate change. Traditional methods for assessing green canopy coverage – the area covered by tree crowns – have often been slow, labor-intensive, and prone to human error. However, a new study introduces an efficient, AI-powered approach to tackle this challenge, utilizing advanced computer vision techniques applied to aerial imagery.
The research, titled “AI-Powered Urban Green Infrastructure Assessment Through Aerial Imagery of an Industrial Township,” by Anisha Dutta from Tata Steel Limited, focuses on Jamshedpur, an industrial city in India known for its planned layout with wide, tree-lined streets. For such a city, accurately calculating green canopy coverage is crucial for urban planning, environmental monitoring, and mitigating the urban heat island effect.
Leveraging Aerial Data and Cloud Computing
The study begins by addressing the data challenge. Instead of traditional satellite imagery or manual measurements, it uses high-resolution aerial imagery collected by drones. Twelve large TIFF image files, each containing seven spectral bands, were gathered from different sections of Jamshedpur. These files, ranging from 7 GB to 48 GB, required significant processing to be compatible with deep learning models.
The preprocessing pipeline is a key innovation. It involves using the GeoTIFF library to extract georeferencing information and the Zarr library to break down the massive image data into smaller, manageable chunks. Specifically, band 2, chosen for its optimal resolution alignment with the training dataset, is converted into NumPy arrays. These arrays are then further chunked into 640×640 pixel segments, with marginal chunks padded with zero values to maintain consistent input size. These processed chunks, numbering around 25,000 per image, are stored in a Google Cloud Storage (GCS) bucket, enabling efficient parallel processing and scalability.
Two AI Approaches: Segmentation vs. Object Detection
The researchers explored two primary computer vision approaches using Convolutional Neural Networks (CNNs) to estimate green canopy coverage:
1. Segmentation Approach: This method uses a CNN-based segmentation model, specifically a PyTorch-based YOLOv5 segmentation model, trained on a dataset of annotated aerial imagery. The goal is to identify and separate green canopies at a pixel level, creating a mask that highlights the exact shape and boundaries of the canopy cover. This provides a detailed, pixel-level understanding of green areas.
2. Object Detection Approach: This approach also uses a PyTorch-based YOLOv5 object detection model. Instead of segmenting pixels, it focuses on identifying individual trees and drawing bounding boxes around them. This method locates and classifies trees within an image.
A comparative study was conducted using an aerial imagery of 23.46 acres of land with validated ground truth data. Both models showed promising results, with the object detection model yielding 44.02% canopy coverage and the segmentation model yielding 44.6%, closely aligning with the validation result of 44.47%. However, the segmentation model was ultimately chosen as the optimal approach. Its ability to provide better pixel-level accuracy, manage overlapping objects more effectively, and align with the naturally continuous and often circular shape of tree canopies made it superior for precise canopy coverage estimation.
Also Read:
- Atlas Urban Index: A New Standard for Urban Development Monitoring with AI and Satellite Imagery
- Deep Learning Unlocks New Methods for Tea Leaf Disease Detection
Overcoming Challenges
The project faced several challenges, including the gigantic size and intricate format of TIFF images, difficulties in transferring large datasets between cloud platforms, a scarcity of ground truth data for training, and the significant computational resources required for CNN-based methods. These were mitigated through the use of GeoTIFF and Zarr libraries, strategic data transfer scheduling, image augmentation techniques, and leveraging Google Colab Enterprise with GCS Bucket for optimized cloud computing.
The final results provide a detailed calculation of green canopy cover for each TIFF file and a visualization of zone-wise green canopy percentages across Jamshedpur. This AI-powered methodology offers valuable insights for urban forestry management, optimizing tree plantation efforts, and assessing the carbon sequestration potential of urban forests. By integrating these insights into sustainable urban planning, cities can foster more resilient, greener, and healthier futures. You can read the full research paper here.
