Land use and land cover (LULC) information has become increasingly important as urban expansion, agricultural development, and environmental degradation continue to reshape landscapes worldwide. Reliable land cover maps help authorities monitor deforestation, evaluate agricultural productivity, manage protected forests, and reduce disaster risks such as flooding and landslides. In Enrekang Regency, where agriculture dominates the landscape and parts of the region have experienced inappropriate land conversion, improving mapping accuracy is particularly valuable for long-term planning and environmental protection.
The researchers focused on Anggeraja District, an area covering approximately 126.96 square kilometers. The region contains forests, plantations, rice fields, dryland agriculture, settlements, rivers, and protected forest areas. Previous studies have reported that thousands of hectares intended for forest protection have been converted into mixed farming and plantation land, highlighting the need for better monitoring tools.
Comparing Two Mapping Approaches
The research compared two widely used image classification approaches:
- Pixel-Based Classification, which analyzes every image pixel individually.
- Object-Based Image Analysis (OBIA), which first groups neighboring pixels into meaningful objects based on their shape, texture, and spatial characteristics before classification.
Each approach was evaluated using two machine learning algorithms:
- Random Forest (RF)
- Support Vector Machine (SVM)
The study utilized freely available Sentinel-2B Level 2A multispectral imagery, downloaded from the European Space Agency (ESA). The classification results were validated using high-resolution SPOT-6/7 imagery and assessed through a confusion matrix, the standard method for measuring classification accuracy in remote sensing.
Object-Based Analysis Consistently Outperformed Pixel-Based Methods
The comparison revealed a clear pattern: incorporating spatial information through object-based analysis significantly improved mapping performance.
The main findings include:
- Object-Based Random Forest achieved approximately 92% overall classification accuracy.
- Object-Based Support Vector Machine achieved approximately 91% accuracy.
- Pixel-Based Random Forest reached approximately 87% accuracy.
- Pixel-Based Support Vector Machine achieved approximately 85% accuracy.
- Random Forest consistently improved overall accuracy by about 1–2% compared with Support Vector Machine under both classification approaches.
The study also found that cloud and cloud shadow classes were classified with nearly perfect accuracy because of their distinctive spectral characteristics. Rice fields, however, remained the most challenging land cover class due to their similarity to other vegetation types.
Another important advantage of Object-Based Image Analysis was its ability to reduce the "salt-and-pepper" effect, a common problem in pixel-based classification where isolated misclassified pixels create noisy maps. By considering groups of pixels instead of individual ones, OBIA generated cleaner and more realistic land cover maps.
Random Forest Was Both More Accurate and More Efficient
Beyond achieving the highest classification accuracy, the Random Forest algorithm also demonstrated greater computational efficiency than Support Vector Machine.
The researchers tested Random Forest models containing between 100 and 400 decision trees. Increasing the number of trees beyond approximately 200–300 produced only marginal improvements, indicating that more complex models do not necessarily generate better classification results. This finding is particularly valuable for organizations processing large satellite datasets because it can reduce computing time without sacrificing accuracy.
The study further showed that the choice of classification algorithm influences not only reported accuracy but also the estimated area assigned to each land cover category. Different algorithms produced different estimates for forests, plantations, rice fields, scrubland, and settlements, demonstrating that methodological decisions can directly affect land management assessments and planning outcomes.
Practical Benefits for Agriculture and Environmental Management
The findings have practical implications for multiple sectors.
More accurate land cover maps can support:
- agricultural expansion planning;
- forest conservation and protected area management;
- regional spatial planning;
- environmental monitoring;
- disaster risk reduction, particularly for floods and landslides;
- land-use policy evaluation; and
- sustainable natural resource management.
Because Sentinel-2 imagery is freely available worldwide, the recommended combination of Object-Based Image Analysis and Random Forest can also be applied in many other agricultural and environmentally sensitive regions beyond Indonesia.
As Nurfadila JS, Rismaneswati, and Syaeful Rahmat from Hasanuddin University conclude, object-based classification consistently provides higher mapping accuracy than pixel-based approaches, while Random Forest delivers the most reliable performance among the evaluated machine learning algorithms. Their findings indicate that selecting the appropriate classification strategy is just as important as selecting the classification algorithm itself when producing dependable land cover maps for agricultural planning and land management.
The researchers recommend that future studies evaluate these methods across different geographic regions, satellite sensors, and time periods, while also exploring advanced machine learning and deep learning techniques to further improve land cover classification performance.
Author Profile
Nurfadila JS is a researcher at Hasanuddin University specializing in remote sensing, Geographic Information Systems (GIS), spatial analysis, and land use/land cover mapping.
Rismaneswati is an academic at Hasanuddin University whose research focuses on geospatial analysis, environmental mapping, and regional planning.
Syaeful Rahmat is a researcher at Hasanuddin University with expertise in machine learning applications for remote sensing, satellite image analysis, and land cover classification.
Source
Article Title: Comparasing Landuse/Landcover Classification Using Pixel and Object Based Random Forest and SVM Algorithm on Sentinel-2 Imagery Over Enrekang Region
Journal: Indonesian Journal of Agriculture and Environmental Analytics (IJAEA)
Publication Year: 2026
Authors: Nurfadila JS, Rismaneswati, Syaeful Rahmat
0 Komentar