Classifying Commodities in GEE

In this step you will import the training polygons which will be used later on in the tutorial.

1. Copy the following code into your new script, (Click the code twice to select all, then hold down Ctrl + C to copy.)
2. Paste this code into your code editor, the code will be highlighted in orange,
3. A pop-up asking whether you would like to convert ‘palm’, ‘cocoa’ … will appear, click “Convert” in the suggestion tooltip.
4. The video below can help you with this step.
5. Afterwards you can close this accordion item and move on to the next one, Setting Parameters and Import Functions.

Set parameters for the processing tool to generate Sentinel-2(L2A) cloud free mosaic. Read more about Sentinel-s (L2A) data.

You can change the map background to satellite view for best visualization effect. To switch from the default Map view, use the buttons in the upper right of the map to select Satellite view.

Set parameters for the processing tool to generate Sentinel-2 cloud free mosaic.

Display the cloud free mosaic that you’ve generated. Different band combination for visualization will emphasize different land features.

Change the bands to the different combinations in order to highlight some of the different properties. The code below shows the possible combinations of bands to highlight specific characteristics.

Due to the similarity of the spectral characteristics of their canopies, some commodities are more difficult to differentiate from natural forest than others (e.g. shaded cocoa). Texture features have proved effective to improve segregating accuracies. Please refer to these two papers for more information on using texture in commodity mapping:

• Numbisi, F. N., Van Coillie, F. and De Wulf, R. (2019). “Delineation of cocoa agroforests using multiseason Sentinel-1 SAR images: a low grey level range reduces uncertainties in GLCM texture-based mapping.” ISPRS International Journal of Geo-Information 8(4): 179.
• Ashiagbor, G., Forkuo, E. K., Asante, W. A., Acheampong, E., Quaye-Ballard, J. A., Boamah, P., Mohammed, Y. and Foli, E. (2020). “Pixel-based and object-oriented approaches in segregating cocoa from forest in the Juabeso-Bia landscape of Ghana.” Remote Sensing Applications: Society and Environment 19: 100349.
  

GEE provides the image.glcm() tool for estimating spatial texture. GLCM computes texture metrics from the Gray Level Co-occurrence Matrix (GLCM) around each pixel of every band. GLCM outputs many other measures of texture. (reference:https://developers.google.com/earth-engine/guides/image_texture).

Sentinel-2 includes a multispectral sensor, however, image.glcm() works on greyscale image. We can either apply image.glcm() on one band (e.g. NIR band), or convert several bands to one grayscale band.

In this tutorial, we will calculate the Luminance from NIR, Green and Blue bands of Sentinel-2 using this equation:

Luminance is shown to performed particularly well for texture recognition and moderately well for object detection and so it is a good bet for GLCM analysis. If you’re eager to learn more, this paper (https://www.ncbi.nlm.nih.gov/pmc/articles/PMC3254613/) assessed the performance of different RGB to Grayscale conversions.

Step 2: apply glcm tool on the reduced grayscale image.
We use kernel size 5 for the analysis and select (CONTRAST layer) from the output layers to represent texture. Many measures of texture are output by image.glcm(). For a complete list of input parameters and output bands. Please note, try different kernel sizes to see if it gives better glcm results.

Display the information. The texture band will be used in the following step as an input band for the classification.

In this tutorial, we will compare the use of 2 different sets of input features:

1. Sentinel-2 spectral bands only
2. Sentinel-2 spectral bands and texture information

Step 0: make a list of band names that we will use as input features in the Random Forest model. All spectral bands from Sentinel-2 are expected to add some information to the classification here.

We will use local contrast (constant_contrast layer) and Difference entropy (constant_dent) from glcm tool to represent spatial texture. Reference for choosing these two indices: Hall-Beyer, M., 2017. Practical guidelines for choosing GLCM textures to use in landscape classification tasks over a range of moderate spatial scales. International Journal of Remote Sensing, 38(5), pp.1312-1338.

We can choose whether or not to include the texture information. This way you can assess how much added value this information has towards the end result.

Scenario 1: Sentinel-2 spectral bands only

Make sure the script between Step 2 (Scenario 2) and step 3 are commented out with the // at the beginning of the line to run scenario 1.

Scenario 2: Sentinel-2 spectral bands and texture information

Comment the two lines of code under ‘Step2 (Scenario 1)’ and uncomment script below to run scenario 2.

A quality training dataset is the key component on which a robust machine-learning model depends. The amount of training data needed depends both on the complexity of your problem and on the complexity of your chosen algorithm. We will show how to collect training data interactively (using opportunity sampling) by drawing polygons using the ‘Geometry drawing tools’ in GEE’s Map Window. This method takes advantage of the very-high resolution satellite imagery that is freely available in GEE as basemap. However, errors may be introduced due to the unknown acquisition date of the background imagery in GEE. A simple cross-validation is used in order to obtain a quick accuracy estimation. A second validation stage using stratified random samples is recommended for improved accuracy assessment. After all, it is important that the date of the imagery that you want to classify is in line with the date of the training data, because you want the land use to be equal in these two sets of imagery. To confirm that these dates are in line, you can use the 3m-resolution Planet Scope data and collect training and/or validation samples for a pre-defined period, Planet is just one option, other data sources can also be used. Find more information on the developer site here . You can also use data from field work or from other high resolution images.

Step 1. Draw polygons for each land use class and save the polygons for each class as a separate geometry layer.

A few examples have been given in this tutorial. You can find them under ‘Geometry imports’ in the upper left of the map. Click on each layer to turn on/off the example layers. The national land use map of Ghana was used as reference for each strata.

Tips for drawing polygons: avoid ambiguous areas in which the land use isn’t clear; avoid edges (edge pixels in your imagery are likely to contain two different types of land use, which confuses the classifier.

Step 2: Merge the geometries of each class (each geometry contains training and validation areas for one strata) as one FeatureCollection.

Step 3: Generate training and validation points within each strata using the sampleRegions function. Each pixel of the input feature image (at 10m scale in this example) that intersects with the training geometry (polygon) is converted to a Feature collection. Each output feature will have one property per band of the input image, as well as the ‘class‘ properties copied from the input feature. Note: In this example we are doing a cross validation…

Some technical details for those interested: Random Forest is an ‘ensemble classifier’, where a group of typically deep, Classification and Regression Tree (CART) classifier probabilities are averaged to predict the most likely class label of a pixel or object. Each tree within the ensemble employs random feature selection and bootstrap sampling with replacement for every split of the CART tree, followed by the averaging of probabilities across all trees, aims to mitigate against high variance. Take a look at some animations to explain this here.

Step 1: Randomly split the sample into training (70%) and validation dataset (30%).

You can uncomment the last line to view the selected training samples.

Step 2: Train a Random Forest classifier. You can tune the model by specifying the parameters (see the Google developers document here. In this example we are using 100 training trees.

Step 3: Validate using an error matrix (note error matrix and confusion matrix are the same thing)

Calculate the validation error matrix and accuracy for both classifiers by using the “errorMatrix” function to generate metrics on the re-substitution accuracy.

From the error matrix we can see that the model has a lower accuracy for class 2 (cocoa). Cocoa is confused with Open Forest (class 3), Palm (class 1) and Other Vegetation.

Step 4: Classify the image and add the map to viewer

Feature importance is used to assess how useful each feature of the model input is in determining the output. This provides us an indication as to which pieces of information we should include in the model, and which ones we could leave out. Again, some technical details for those interested: The Random Forest classifier in Google Earth Engine is based on SMILE, where feature importance is calculated by summing up the decreases for each individual variable (GINI, information gain, etc.) over all trees in the forest. It is a fast way to calculate variable importance and is often very consistent with the permutation importance measure (more information refer to the GitHub page)

We use the classifer.explain() function in GEE to explore the variable importance of the trained Random Forest model.