5  GEE

5.1 Summary

5.1.1 What’s GEE?

• it uses Javascript, which is website programming language, variables are defined by var

• it allows massive datasets for quickly planetary scale analysis.

• in GEE, image refers to raster, feature refers to vector, ImageCollection refers to image stack, and FeatureCollection refers to feature stack(lots of polygons)

• code runs on the client side, which is the browser, while the data is stored on the server side(ee)

5.1.2 Why GEE?

• Scale: The scale is set by the output instead of input.

  • When doing analysis, GEE would aggregate the image to fit a 256*256 grid, and they would also select the pyramid with the closest scale and resample using nearest neighbor by default.

• Projection: This is what we don’t need to worry about in GEE. It converts all data into Mercator projection when displaying

5.1.3 The Usages of GEE

5.1.3.1 Building Blocks

• objects: vector, raster, feature, string, number

  • each belongs to a class

  • each class has specific GEE functions/methods

• geometry: point/line/(multi)polygon without attributes

• feature: geometry with attributes

  • feature collection: multiple features with attributes

5.1.3.2 Typical processes

• geometry operations: joins, zonal statistics, filtering of images or specific values

• methods: machine learning, supervised and unsupervised classification, etc.

• applications/outputs: online charts, scalable geospatial applications with GEE

• Instances:

  • Reducing images: reduce the collection to the extreme values for each pixel(collection.reduce(ee.Reducer()))

    • By region: most useful for zonal statistics(reduceRegion(), image.reduceRegions())

    • By neibourhood: a window of pixels surrounding a central pixel

  • Linear regression: see the change over time in pixel values(linearFit())

  • Joins: first put image collections or feature collections in a filter(ee.Filter), then join with join.apply().

5.2 Application

In addition to a large repository of raw remotely sensed imagery, the GEE data catalog also offers preprocessed, mosaic images that have been processed to remove clouds. Table 1, as listed by Tamiminia et al. (2020a), provides a more comprehensive overview of the capabilities of packages in the GEE platform, I think, which would be enlightening for us to explore. Still, I believe that the most effective way of learning these things would be open the browser and look for the GEE APIs. And try to write your own code with interesting datasets.

They reviewed 349 papers used GEE in a wide variety of remote sensing applications, and categorized them into 11 different groups(shown in Figure below). Different values in the gear show the number of papers in related areas. It has shown that most studies are on crop mapping, including vegetation, rice paddy, and agricultural monitoring. This is followed by studies on water, land cover/land use, and disaster. The least number of studies are on data processing.

It is interesting to note that they also categorized all studies into two groups based on the type of data used: remote sensing data and ready-to-use products(such as NDVI and land cover). These two main categories were further subdivided into smaller categories(shown in Figure below) based on the methods. The most used methods are machine learning, while a small part of research focus on time-series analysis, feature extraction, image composite-visual interpretation and image pre-processing.

In machine learning, classification, clustering, regression and dimension deduction are four main algorithms(Holloway and Mengersen 2018). While in the papers Tamiminia et al. (2020b) investigated, they find that in GEE application classification and regression are the most popular machine learning algorithms in remote sensing. In detail, DT, CART, KNN, non-linear SVM, RF, and ANN are the most often-use non-parametric classification algorithms. In the category of others, time-series analysis, feature extraction, image composite-visual interpretation and image processing were applied to satellite images.

Even GEE is powerful for geospatial and big data analysis, it has several limitations. Firstly, despite it has an extensive archive of satellite images, it has limited historical and high-resolution data. Notably, it could be challenging to implement new algorithms. For instance, Deep Learning has becoming extensively used in image classification. However, DL algorithms are not yet supported directly by GEE. The open-source framework like Tensorflow has been an important way of using DL in GEE.

5.3 Reflection

• As shown in practical, while Google Earth Engine (GEE) is a powerful platform for remote sensing and image analysis, it can be complex and verbose when compared to other software packages like R or QGIS, especially when performing simple tasks like masking. I think, it may because GEE is primarily designed for image analysis rather than vector-based processing.

• Besides, its programming language, which is based on JavaScript, can be challenging for me. I have discovered a Python library called geemap, which I find more user-friendly than using JavaScript directly in GEE. Since I am more comfortable with Python, I plan to use geemap in my future studies to explore remote sensing and image analysis tasks further.

Holloway, Jacinta, and Kerrie Mengersen. 2018. “Statistical Machine Learning Methods and Remote Sensing for Sustainable Development Goals: A Review.” Remote Sensing 10 (9): 1365. https://doi.org/10.3390/rs10091365.

Tamiminia, Haifa, Bahram Salehi, Masoud Mahdianpari, Lindi Quackenbush, Sarina Adeli, and Brian Brisco. 2020a. “Google Earth Engine for Geo-Big Data Applications: A Meta-Analysis and Systematic Review.” ISPRS Journal of Photogrammetry and Remote Sensing 164 (June): 152–70. https://doi.org/10.1016/j.isprsjprs.2020.04.001.

———. 2020b. “Google Earth Engine for Geo-Big Data Applications: A Meta-Analysis and Systematic Review.” ISPRS Journal of Photogrammetry and Remote Sensing 164 (June): 152–70. https://doi.org/10.1016/j.isprsjprs.2020.04.001.