DefinePK

Research Objective

To compare the effectiveness of pixel-based and object-based approaches for classifying crop types (cotton, rice, and sugarcane) using machine learning algorithms on integrated Sentinel-1 and Sentinel-2 satellite data.

Methodology

The study utilized Synthetic Aperture Radar (SAR) Sentinel-1 and optical Sentinel-2 satellite data processed on the Google Earth Engine (GEE) platform. Pre-processing involved speckle filtering, border noise removal, radiometric and geometric corrections, and the computation of vegetation indices (NDVI, NDWI, NDMI, BSI). Composite images were created for specific crop growth periods (August for cotton, September for rice and sugarcane). Two object-based segmentation approaches were employed: Simple Non-Iterative Clustering (SNIC) in GEE and Multi-Resolution Segmentation in e-Cognition. The Random Forest (RF) algorithm was applied to both pixel-based and object-based approaches. Field sample data collected in Rahim Yar Khan, Pakistan, was used for classification and validation. Accuracy was assessed using confusion matrices and kappa coefficients.

Methodology Flowchart

                        graph TD
    A[Acquire Sentinel-1 & Sentinel-2 Data] --> B[Pre-process Data on GEE];
    B --> C[Compute Vegetation Indices];
    C --> D[Create Composite Images];
    D --> E[Apply Segmentation];
        E -- SNIC --> F1[SNIC Segmentation GEE];
        E -- OBIA --> F2[OBIA Segmentation e-Cognition];
    F1 --> G[Apply Random Forest Classification];
    F2 --> G;
    G --> H[Validate Classification];
    H --> I[Assess Accuracy];
    I --> J[Compare Approaches];                    

Discussion

The study highlights that object-based approaches, by analyzing groups of pixels, reduce noise and improve classification accuracy compared to traditional pixel-based methods, especially in heterogeneous agricultural landscapes. The SNIC algorithm in GEE showed superior performance, suggesting the importance of segmentation parameters and processing platforms. The integration of Sentinel-1 and Sentinel-2 data proved beneficial due to their complementary strengths. Limitations included potential misclassification due to mixed pixels and crop growth variability.

Key Findings

The pixel-based Random Forest classification achieved the highest overall accuracy (98%) and a kappa coefficient of 92%. The object-based approach using SNIC in GEE achieved 96% accuracy with a kappa coefficient of 95%. The OBIA approach in e-Cognition attained 89% accuracy. While the pixel-based method showed high overall accuracy, the object-based methods, particularly SNIC in GEE, demonstrated better producer and user accuracy, indicating improved delineation of crop boundaries.

Conclusion

Integrating Sentinel-1, Sentinel-2, and derived indices is advantageous for both pixel-based and object-based crop classification. The object-based approach using SNIC in GEE outperformed pixel-based methods and OBIA in e-Cognition, likely due to better handling of spatial information and noise reduction. The choice of classification method should consider the specific characteristics of the study area, and further refinement of object-based methods is recommended for complex agricultural environments.

Fact Check

1. Accuracy of Pixel-Based RF: The study reports 98% accuracy for the pixel-based Random Forest classification.
2. Kappa Coefficient for SNIC: The SNIC object-based approach achieved a kappa coefficient of 95%.
3. Study Area: The research was conducted in Rahim Yar Khan, Pakistan.

Mind Map