Landsat 8 Classification Map
Define ROI
in this tutorial, which speckle???
Land cover monitoring using remotely sensed data requires robust classification methods which allow for the accurate mapping of complex land cover and land use categories. Random forest (RF) is a powerful machine learning classifier. https://www.academia.edu/2052100/An_assessment_of_the_effectiveness_of_a_random_forest_classifier_for_land_cover_classification Random forest is a tree-based algorithm. Its base learner is the decision tree. Therefore, a random forest can be considered as an ensemble (group) of decision trees. https://towardsdatascience.com/how-to-train-a-regression-model-using-a-random-forest-c1cf16288f6b The algorithm requires training data for making predictions. A training dataset comprises observations and features representing a selected land cover type. The decision trees produce different outputs, depending on the training data fed to the random forest algorithm. These outputs will be ranked, and the highest will be selected as the final output. https://www.section.io/engineering-education/introduction-to-random-forest-in-machine-learning/#:~:text=A%20training%20dataset%20comprises%20observations,selected%20as%20the%20final%20output. We will use SAR backscatter values to define the land cover classes. The backscatter intensity should increase from bare_fields to forest from dark to light. , and collect training data for each class to ingest into the Random Forest.
Sentinel -1 Landsat 8
image slider
slide between images with the circle
Random Forest Classifier
L 8 Classification
SAR & Optical Classification image slider
switch between images with the diamonds
Sentinel -1 Landsat 8
image slider
L 8 RGB
S-1 VH
S-1 VH (filtered)
Landsat 8 RGB (R: band 4, G: band 3, B: band 1)
Filtered SAR image slider
L 8 NDVI
Land cover monitoring using remotely sensed data requires robust classification methods which allow for the accurate mapping of complex land cover and land use categories. Random forest (RF) is a powerful machine learning classifier. https://www.academia.edu/2052100/An_assessment_of_the_effectiveness_of_a_random_forest_classifier_for_land_cover_classification Random forest is a tree-based algorithm. Its base learner is the decision tree. Therefore, a random forest can be considered as an ensemble (group) of decision trees. https://towardsdatascience.com/how-to-train-a-regression-model-using-a-random-forest-c1cf16288f6b The algorithm requires training data for making predictions. A training dataset comprises observations and features representing a selected land cover type. The decision trees produce different outputs, depending on the training data fed to the random forest algorithm. These outputs will be ranked, and the highest will be selected as the final output. https://www.section.io/engineering-education/introduction-to-random-forest-in-machine-learning/#:~:text=A%20training%20dataset%20comprises%20observations,selected%20as%20the%20final%20output. We will use SAR backscatter values to define the land cover classes. The backscatter intensity should increase from bare_fields to forest from dark to light. , and collect training data for each class to ingest into the Random Forest.
S-1 Classification
Sentinel-1 VH
Random Forest classification
Collecting Training Data for Random Forest Classifier
Land Cover Map (SAR & Optical Classified Image)
S-1 VV
Landsat 8 RGB (R: band 4, G: band 3, B: band 1)
slide between images with the circle
NDVI image slider
Sentinel-1 VV
Sentinel-1 VH - Speckle Filtered
(50 m smoothing radious)
Load and Subset the EO Data
Land Cover Map
image slider
Filter Speckle
SAR & Optical Classification image slider
how to calculate accuracy assessment
Accuracy Assessment
Mask Cloud & Cloud Shadow
We will create a Land Cover Map based on EO data from both:
Optical: Landsat 8
Radar: Sentinel-1 sensors.
Load the image collections and crop by the ROI to reduce the processing efforts.
We will map the land cover for the beginning of August 1019, to do this filter the image collection to take only images from 01.08 to 10.08.2019.
SAR - Random Forest
Classification
Accuracy Assessment
Collect Training Data
S-1 VH filtered
XXXXXXX Classification SAR & Optical
z
Load and Prepare the EO Data
Calculate NDVI
Speckle Filtering of SAR Data
Define Region of Interest
S-1 & L 8 Land Cover Map
XXXXX
Optical - Random Forest
Classification
- Every grey vector corresponds to a scatterer in the resolution cell.
- Resultant amplitute of the pixel (red vector) is the coherent sum of all those individual contributions
To learn more check the
NDVI image slider
Xxxxxxx
Filtered SAR image slider
Final Accuracy
The exclusion of clouds and cloud shadows is an important processing step that is usually done in an early pre-processing stage of the optical data.
In this tutorial you will use the ‘QA_PIXEL’ bitmask in the Landsat 8 data to mask out the clouds and their shadows.
resource on Speckle at EO College
Masking of Cloud and Cloud Shadow
on the Optical Images
Speckle is an inherently exist salt-pepper effect that degreades the SAR images and makes interpretation of features difficult by reducing the effectiveness of image segmentation and classification. Therefore many approaches exist to reduce the speckle effect.
Speckle is related to the radar image acquisition principles and happens due to the interference of waves reflected from many elementary scatterers.
Land Cover Map image slider
Classification with
SAR & Optical
Normalized Difference Vegetation Index (NDVI) quantifies vegetation by measuring the difference between near-infrared (which vegetation strongly reflects) and red light (which vegetation absorbs).”
What is NDVI (Normalized Difference Vegetation Index)? NDVI always ranges from -1 to +1. But there isn’t a distinct boundary for each type of land cover. For example, when you have negative values, it’s highly likely that it’s water. On the other hand, if you have an NDVI value close to +1, there’s a high possibility that it’s dense green leaves. But when NDVI is close to zero, there are likely no green leaves and it could even be an urbanized area. NDVI is the most common index that analysts use in remote sensing.
But how do you calculate it? What do NDVI values represent? How do Earth scientists use NDVI? Devami: https://gisgeography.com/ndvi-normalized-difference-vegetation-index/ We will use NDVI as information for the classification of land covers.
Generating NDVI Images
Accuracy Assessment
Sentinel-1 VH
Landsat 8 NDVI
Landsat 8 RGB (R: band 4, G: band 3, B: band 1)
Sentinel-1 VH (speckle filtered)
Sentinel-1 Classification Map
Land Cover Map
image slider
SAR & Optical Classification image slider
Filter Images for Summer Dates
Load and Prepare the Data
Preprocess for Classification
Specify the Bands for Classification
Split samples for Training and Testing
Masked Image
Assess Accuracy
Create Land Cover Map & Export
Combine Land Cover Classes, specify the bands and training points, and create test and training dataset.
1. Collect Training Data for LC Classes: In this example, we use Land Cover Classes:
- Coniferous
- Mixed Forest
- Diciduous
- Cultivated
- Water
And the masked Classes:
- impervious layer
- Cloud
You can use the feature Classes provided by the tutorial or create your own.Click here for details.
2. Specify the Bands for Classification: Select the Bands of the Landsat 8 data that are sensitive to the vegetation, these are:
B3 Green
B4 Red
B5 near-infrared
B6 shortwave infrared 1
B7 shortwave infrared 2
3. Split Samples for Training and Testing: At this step, we are randomly splitting the samples (collected training data) to set some aside for testing the model’s accuracy. Roughly 80% for training, 20% for testing.
Final Land Cover Classification Map
Collect training data
In this tutorial, we are interested in identifying and classifying the developed plants, therefore, we masked the land cover type that is not relevant to this objective, i.e clouds and urban areas.
StartFragmentThe exclusion of clouds and cloud shadows is an important processing step that is usually done in an early pre-processing stage of the optical data. In this tutorial, you will use the 'QA PIXEL’ (QA stands for Quality Assessment) bitmask in the Landsat 8 data to mask out the clouds (Bit 5) and their shadows (Bit 3).EndFragment
For the exclusion of the urban areas, at this step, you will import the USGS’ National Land Cover Database (NLCD) directly from GEE. Here, you will use the ‘impervious layer’ of the NLCD data to amsk out ubran areas.
Note that, this is an example of how you can use pre-existing layers in GEE. Depending on your specific application, you can use your own data to mask the features that are not in your interest.
Add existing surface Layer and mask out clouds & urban areas
RGB Image
Mask the Clouds & Urban Areas
Predict the land use using random forest and the training data.
Random Forest (RF) Classification learns from training data and identifies statistical patterns in large datasets.
- it is a tree-based machine-learning algorithm that uses a series of decision trees to select the best classification for all pixels within the imagery.
- The algorithm iterates over decision trees that allow voting for the best solution.
Advantages of RF:
- Use of multiple trees reduces the risk of overfitting
- Training time is shorter and not sensitive to outliers in training data
- Runs efficiently and produces high accuracy for large datasets
- Easy to parameterise
Limitations of RF:
- Algorithm cannot predict spectral range beyond training data
- Training data must capture the entire spectral range
Classify with Random Forest
Land Cover Map
Land Cover Map
Assess Accuracy
Classify with Random Forest
Preprocess for Classification
Create a Legend (assign appropriate colors to each class), Display the final Land Cover Classification, and provide Export Options.
RGB Image
Land Cover Image
Create Land Cover Map & Export
Masked Image
Create Land Cover Map & Export
Use the test data to evaluate the classification
- Accuracy refers to the degree of corrospondence between classification and reality.
- Accuracy Assessment involves the comparison of the image classification to reference data.
- References can include ground data or a subset of training points withheld for accuracy assessment purposes.
- Comparison of reference data and classifications is typically done using a confusion (or error) matrix to compile these comparisons.
Error matrix is a table of references to predicted classes.
Here is an example of Error matrix
display in the GEE console for
accuracy assessment. The number of
correctly classified pixels is shown
along the diagonal of the error matrix.
For instance, in the example from GEE,
the correclty classified number
of pixels for class4, is 1107.
The overall accuracy based
on this error matrix is then 0.99111.
Filtered Composite of the study area
Add the imagery, filter to the area of interest and date range and make a composite.
1. Load the optical image: Insert the optical image collection (Landsat 8) and filter by area using an imported shape file. The shapefile is provided for this tutorial but you can also create geometry on the GEE map view.
2. Filter Images for Summer Dates: The vegetation cover at its peak during the summer month. In winter, the state of vegetation is different, for instance, the leaves of the trees fall, or snow might cover the vegetation. Therefore we filter the image collection to summer dates only.
Masked Image
Land Cover Map
Mask the Clouds & Urban Areas
Load the optical image
Land Cover Map
