DefinePK

Research Objective

To analyze and compare the performance of three machine learning methods (Maximum Likelihood Classification (MLC), Support Vector Machines (SVM), and Random Forest (RF)) for land cover mapping using Landsat 8 multispectral imagery, PCA-derived images, and spectral indices, and to validate the most accurate map against the European Space Agency's World Cover dataset.

Methodology

The study involved acquiring Landsat 8 OLI satellite imagery and the ESA's World Cover LULC map. Landsat 8 imagery was pre-processed, and Principal Component Analysis (PCA) was applied for dimensionality reduction. Spectral indices (NDVI, NDSI, NDBI) were calculated. Three supervised classification algorithms (MLC, SVM, RF) were applied to multispectral images, PCA images, and stacked indices. A total of 530 sample points were used (350 for training, 180 for validation). Accuracy assessment was performed using a confusion matrix, calculating overall accuracy (OA), user's accuracy (UA), producer's accuracy (PA), and the Kappa coefficient (Kc). The ESA World Cover LULC product was used for validation.

Methodology Flowchart

                        graph TD
    A[Acquire Landsat 8 Imagery and ESA World Cover Map] --> B[Pre-process Landsat 8 Imagery];
    B --> C[Apply PCA and Calculate Band Indices];
    C --> D[Prepare Datasets: Multispectral, PCA, Stacked Indices];
    D --> E[Train and Apply MLC, SVM, RF Classifiers];
    E --> F[Assess Classification Accuracy: OA, Kappa, UA, PA];
    F --> G[Validate Results with ESA World Cover Map];
    G --> H[Compare Algorithm Performance and Data Types];
    H --> I[Draw Conclusions and Discuss Findings];                    

Discussion

SVM's superior performance is attributed to its robustness and ability to handle complex linkages in data. PCA enhanced classification accuracy by reducing dimensionality and retaining variance. Stacked indices, particularly NDBI, sometimes led to misclassifications, impacting overall accuracy. The study highlights that classifier performance is influenced by both the algorithm and the input data. The lower accuracy of the ESA World Cover map is attributed to temporal and resolution discrepancies.

Key Findings

The SVM algorithm outperformed MLC and RF, achieving the highest overall accuracy of 90.43% with a kappa coefficient of 0.8792 when applied to the Landsat 8 multispectral image. PCA-derived images generally improved classification accuracy across all methods compared to stacked indices. The ESA's World Cover LULC product showed an overall accuracy of 70.67% in the study area. Snow cover was the most accurately classified class across all methods.

Conclusion

Machine learning algorithms, particularly SVM, are effective for LULC classification using Landsat data and its derivatives. SVM demonstrated the highest accuracy across all tested datasets. The Landsat 8 multispectral image provided the best results. The study provides valuable insights for LULC mapping in mountainous regions and for policymakers.

Fact Check

* The SVM algorithm achieved an overall accuracy of 90.43% and a kappa coefficient of 0.8792.
* The ESA's World Cover LULC product achieved an overall accuracy of 70.67% in the study area.
* Landsat 8 OLI satellite imagery has a 30-meter spatial resolution.

Mind Map