Summary of report
To investigate the efficiency of Sentinel-2 data and Landsat-8 data in uncovering data insights, our team conducted a comparison analysis of land cover classification between the two data sources. By analyzing different classification results with various machine learning algorithms, we can get a deep understanding of discriminating machine learning problems.
In this study, we have chosen North of Singapore as our area of focus due to its wide variety of classification classes. The following satellite imagery and bands were selected for analysis.
| Landsat 8 | Sentinel-2 |
|---|---|
| 20180524 | 20200126 |
| Bands 1 - 7 (30m resolution) | Bands 2,3,4,8 (10m resolution) |
| Low Cloud Cover Observed | Low Cloud Cover Observed |
| CRS WGS_84 | CRS WGS_84 |
Through utilizing the band sets identified to perform different combinations of color-composite, we have defined the following 9 classes to use as our training sites.
Water Body
Inland water
Bare Land
Natural Vegetation
Managed Vegetation
Built-up 1 (High density)
Built-up 2 (High density)
Built-up 3(Low density)
Impervious surfaces
We then used SCP to define areas that would used as training sites and generated the ROIs (Regions of Interest) and spectral signatures required for the classification of a band set. Through a series of comparisons, we found out maximum likelihood analysis would allow us to yield higher accuracy than minimum distance analysis.
Results discussed in the following table.
| Landsat 8 | Sentinel-2 | |
|---|---|---|
| Overall Accuracy | 51.24% | 43.69% |
| Kappa Hat Classification | 0.3348 | 0.4216 |
| Water Bodies | Identified water bodies from inland land | Identify only inland water |
| Vegetation | Identify only managed vegetation | Identified natural vegetation from managed vegetation |
| Bare Land | Accurate | Misclassification |
| Spatial Resolution | Lower | Higher - clearer images |
Both satellite imagery has their own benefits and shortcomings. We hope that the results would be useful for future land cover classification and in our future works, we aim to merge Landsat-8 and Sentinel-2 imagery to improve the overall accuracy.
Word Count: 306