Adaptive Land-Use Change Detection Using Multi-Temporal High-Resolution Satellite Imagery and Deep Learning in a GIS Framework

Project Abstract

This study presents a novel framework for adaptive land-use change detection by synergistically integrating multi-temporal high-resolution satellite imagery with deep learning within a geographic information system (GIS) environment. The core objective is to enhance spatiotemporal accuracy, robustness, and scalability of land-use mapping under dynamic urban and peri-urban transformations. We leverage a curated dataset consisting of multi-spectral, high-resolution images captured over a ten-year horizon, incorporating diverse land-cover classes including residential, commercial, industrial, agricultural, and green spaces. A multi-branch deep neural network architecture is developed to capture both pixel-level texture cues and object-level semantic context, using temporal fusion strategies that integrate feature representations across time via attention-based modules and recurrent mechanisms. To address class imbalance and sparse change events, the model incorporates adaptive loss functions and focal regularization informed by regional land-change statistics derived from ancillary census and cadastral data. The pipeline includes rigorous pre-processing steps for radiometric normalization, geometric co-registration, and cloud/mask filtering, followed by change-aware feature extraction that combines spectral indices, texture descriptors, and LiDAR-derived height information where available. Change detection is cast as a conditional segmentation problem, enabling precise delineation of newly developed parcels, demolished structures, and land-cover transitions, while maintaining consistency with existing GIS layers through spatial co-location constraints. The framework is designed to be adaptive, allowing continuous learning from newly acquired imagery via incremental training and domain adaptation to different geographic regions with minimal labeling effort. Evaluation is conducted across multiple case study sites representing heterogeneous landscapes, including rapidly urbanizing cities, peri-urban fringes, and agricultural hinterlands. Metrics such as overall accuracy, per-class F1-score, kappa, and change-level metrics (change detection rate, false alarm rate) are complemented by area-weighted accuracy and transition accuracy to capture both immediacy and persistence of detected changes. Results demonstrate superior performance over traditional pixel-based and conventional machine learning approaches, particularly in complex edge cases where small-scale developments and fragmented land parcels challenge routine classifiers. The study also analyzes computational efficiency, proposing pruning strategies and hardware-accelerated inference to enable near-real-time monitoring within GIS platforms. Furthermore, the research explores interpretability through gradient-based saliency maps and feature importance analyses to elucidate the drivers of detected changes, supporting policy-makers and urban planners in decision-making, zoning compliance, and sustainable land management. Finally, the proposed adaptive framework provides a scalable blueprint for integrating multi-temporal high-resolution remote sensing data with advanced deep learning techniques, delivering timely, accurate, and actionable land-use change information in a GIS context to support resilient urban planning and environmental stewardship.

Project Overview

What This Project Is About

---

A beginner-friendly overview of using satellite images taken at different times to automatically detect how land use changes, like from farmland to housing, by teaching a computer to learn from examples. The project combines map data, imagery, and simple methods to identify where and when changes happen.

---

---

The Problem It Addresses

---

Many areas change quickly due to urban growth, agriculture, or deforestation, and traditional methods to track these changes are slow and manual. This project aims to create faster, more accurate methods that can automatically spot changes over time to help planners and researchers.

---

---

Objectives of the Project

---

1. Understand how land use changes can be spotted in photos taken at different times.
2. Build a simple pipeline that combines high-resolution images with an easy-to-train learning model.
3. Evaluate how well the method detects changes compared with existing approaches.
4. Provide a clear, usable workflow that can be applied to new areas.

---

---

What You Will Do Step by Step

---

1) Gather multi-temporal high-resolution images of a study area. 2) Prepare the data by aligning images and labeling examples of different land uses. 3) Train a straightforward deep learning model to recognize land-use types and changes. 4) Test the model on new images and measure accuracy. 5) Analyze results, discuss strengths and limits, and suggest improvements.

---

---

Expected Outcome

A simple, repeatable method to detect land-use changes over time with high accuracy, plus a clear guide on applying it to other regions. The project should produce a ready-to-use workflow and insights for urban planning and environmental monitoring.