[Paper Review] Rule based End-to-End Learning Framework for Urban Growth Prediction

Saptarshi Pal, Soumya K. Ghosh|arXiv (Cornell University)|Nov 29, 2017

Land Use and Ecosystem Services18 references3 citations

TL;DR

This paper proposes a novel end-to-end learning framework that directly extracts cellular automata rules from Landsat satellite imagery for urban growth prediction, bypassing manual feature engineering. By leveraging decision trees and representation learning, the framework achieves superior performance—outperforming existing methods in Figure of Merit, Producer’s and User’s accuracy, and overall accuracy—while automatically capturing spatial constraints like water bodies and swamps without explicit input.

ABSTRACT

Due to the rapid growth of urban areas in the past decades, it has become increasingly important to model and monitor urban growth in mega cities. Although several researchers have proposed models for simulating urban growth, they have been primarily dependent on various manually selected spatial and nonspatial explanatory features for building models. A practical difficulty with this approach is manual selection procedure, which tends to make model design process laborious and non-generic. Despite the fact that explanatory features provide us with the understanding of a complex process, there has been no standard set of features in urban growth prediction over which scholars have consensus. Hence, design and deploying of systems for urban growth prediction have remained challenging tasks. In order to reduce the dependency on human devised features, we have proposed a novel End-to-End prediction framework to represent remotely sensed satellite data in terms of rules of a cellular automata model in order to improve the performance of urban growth prediction. Using our End-to-End framework, we have achieved superior performance in Figure of Merit, Producer's accuracy, User's accuracy, and Overall accuracy metrics respectively over existing learning based methods.

Motivation & Objective

• Address the labor-intensive and non-generic nature of manual feature engineering in urban growth modeling.
• Reduce dependency on manually selected spatial and spectral features for urban growth prediction.
• Develop an end-to-end learning framework that directly learns rules from raw satellite imagery for improved generalization and performance.
• Enable automatic discovery of spatial constraints (e.g., water bodies, swamps) without explicit labeling or feature engineering.
• Improve prediction accuracy and robustness using representation learning on spectral and spatial data from Landsat imagery.

Proposed method

• Use raw Landsat satellite imagery as input to bypass traditional feature engineering steps.
• Apply end-to-end learning to map spectral and spatial patterns directly to cellular automata transition rules.
• Employ decision trees and random forests as interpretable knowledge structures to extract and represent learned rules.
• Train models using a single-layer MLP, SGD, Naive Bayes, KNN, AdaBoost, and other classifiers for comparative evaluation.
• Utilize representation learning to automatically capture complex spatial and spectral patterns from remote sensing data.
• Incorporate temporal data across multiple years (1991–2011) to model urban transition dynamics and predict future growth (2051).

Experimental results

Research questions

• RQ1 Can an end-to-end learning framework eliminate the need for manual feature engineering in urban growth prediction?
• RQ2 How does direct rule learning from satellite imagery compare to traditional methods using engineered spatial and spectral features?
• RQ3 To what extent can the framework automatically learn and respect natural barriers such as water bodies and swamps without explicit input?
• RQ4 What is the performance of the end-to-end framework relative to established models like logistic regression, SVM, and random forests?
• RQ5 Can the framework generalize across different urban regions and maintain high accuracy despite data imbalance and resolution constraints?

Key findings

• The end-to-end framework achieved a Figure of Merit (FoM) of 0.911805 with a single-layer MLP, outperforming existing methods including logistic regression, SVM, and random forests.
• The framework achieved 91.18% overall accuracy using a single-layer MLP and 90.08% with random forests, demonstrating strong predictive performance.
• Producer’s accuracy and User’s accuracy were significantly higher than those of traditional models, indicating improved detection and classification of built-up areas.
• The model successfully avoided urban encroachment into water bodies and swamps in 2051 predictions, even without explicit input about these features.
• The decision tree-based implementation showed minimal overfitting and better generalization compared to other end-to-end models like KNN and Naive Bayes.
• The framework demonstrated robustness to data imbalance and could learn transition patterns despite reduced temporal resolution and fixed spatial resolution.