TLDR: A research paper by Gurjard et al. uses remote sensing and machine learning (CNNs, Random Forest, ANN, K-means) to analyze land cover changes in Nadi, Fiji, from 2013 to 2024. The study reveals rapid urbanization, with urban areas expanding and replacing agricultural land and mangroves. Convolutional Neural Networks (CNNs) proved most accurate for classification, providing crucial insights for land use planning in the rapidly developing region.
Fiji, a vibrant developing nation, is experiencing significant changes to its landscape, driven by rapid urbanization and large-scale development projects like new housing, roads, and civil infrastructure. A recent study delves into these transformations, specifically focusing on land use and land cover changes in Nadi, Western Fiji, between 2013 and 2024.
The research, titled “Landcover classification and change detection using remote sensing and machine learning: a case study of Western Fiji,” was conducted by Yadvendra Gurjard, Ruoni Wena, Ehsan Farahbakhshb, and Rohitash Chandraa. Their primary goal was to offer technical insights into land cover and land use modeling, as well as change detection, to support sustainable development in the region.
Unveiling Changes from Space
To achieve this, the researchers employed a powerful combination of remote sensing technology and advanced machine learning techniques. They utilized images from the Landsat-8 satellite, which provides multi-spectral data of the Earth’s surface. A crucial step involved creating a detailed training dataset with labels for supervised machine learning, essentially teaching the computer what different types of land look like.
The study leveraged Google Earth Engine, a cloud-based platform, for processing satellite imagery. This platform was instrumental in generating land cover maps using an unsupervised machine learning method called k-means clustering. Additionally, convolutional neural networks (CNNs), a type of deep learning model particularly effective for image processing, were used to classify specific land cover types within selected regions. The final output included visualizations that clearly highlighted urban area changes over time, allowing for effective monitoring of these shifts.
The Methodology Behind the Maps
The process began with meticulous data preparation. Landsat 8 Operational Land Imager (OLI) data, with its 30-meter spatial resolution, was used. To ensure clear images, a sophisticated cloud removal process was implemented using Google Earth Engine, creating median composite images for each year from 2013 to 2022 to minimize atmospheric disturbances. To further enhance classification, three key remote sensing indices were calculated: the Normalized Difference Vegetation Index (NDVI) for vegetation, the Normalized Difference Water Index (NDWI) and Modified NDWI (MNDWI) for water bodies, and the Normalized Difference Built-up Index (NDBI) for urban areas.
The classification scheme adopted for the study was a modified version of the European Union Land Use/Cover Area Frame Survey (LUCAS) scheme. It categorized land into seven major classes: Urban Areas, Grass/Agricultural Land, Forest, Bare Soil, Water Bodies, Coastal Areas, and Wetland. Manual labeling of ground-truth data, using high-resolution historical images from Google Earth, was a critical step in training the supervised machine learning models. The researchers also carefully determined an appropriate sample size for training, finding that approximately 0.25% of the total pixels yielded the most effective results for capturing the diversity of land cover types.
Comparing Machine Learning Approaches
The study compared the performance of several machine learning models: k-means clustering (unsupervised), Random Forest, Multilayer Perceptron (a simple neural network), and Convolutional Neural Networks (CNNs) (all supervised). The models were evaluated based on accuracy, precision, recall, and F1 score. While all supervised models showed high accuracy, the CNN model consistently outperformed the others, achieving an accuracy of 99.05%. This superior performance was attributed to CNN’s ability to effectively learn spatial context information from the image data.
The k-means unsupervised method, while useful, tended to overestimate urban areas, often misclassifying burnt bare soils and red soils in mountain regions as urban. The supervised models, especially CNN, were more adept at distinguishing between similar land cover types and accurately classifying urban areas, which was crucial for the change detection aspect of the study.
Nadi’s Evolving Landscape
Using the best-performing CNN model, the researchers generated land cover maps for Nadi from 2013 to 2024. These maps revealed a clear pattern of rapid urbanization. An urban expansion map showed that urban areas grew outwards along existing peripheries. A replacement map further illustrated that this growth in Nadi downtown, particularly near the coastline, corresponded to a significant reduction in grass and agricultural land. Notably, the study found that mangrove areas in Nadi Bay have been gradually replaced by urban development since 2016.
The study acknowledges certain challenges, such as potential human errors in manual labeling due to the use of median composite images and the limited availability of official ground truth data for direct comparison. However, the high assessment scores and the detailed change detection maps provide valuable insights into Nadi’s evolving landscape.
Also Read:
- AI Maps City Dwellers: Deep Learning Estimates Gandhinagar’s Population with Satellite Imagery
- Boosting Lightweight AI for Satellite Hyperspectral Data
Looking Ahead
This research provides a robust framework for monitoring land cover changes in Fiji and other developing regions facing similar urbanization pressures. The findings underscore the effectiveness of integrating remote sensing with advanced machine learning, particularly CNNs, for accurate land cover classification and change detection. The rapid urbanization observed in Nadi highlights the need for continued monitoring and informed planning to manage the environmental and societal impacts of such growth. For more details, you can read the full research paper here.
