The main file used in this project is SupervisedClassification.js. Here I show how to integrate information on crop cycles with a CART classification algorithm in Google Earth Engine to identify the annual extent of sugarcane in Pakistan over the course of 34 years.
There are 5 broad steps to this process:
-
Generating training data. I identify sugarcane using crop masks provided by SUPARCO (the Pakistani Space Agency), and other types of landcover (urban, water, desert) visually.
-
Selecting imagery based on crop cycles. This was done by first filtering the imagery to 2015 (the year the crop masks were created), and to months in which sugarcane is in peak growth while other local crops (mostly rice and wheat) are not.
-
Spectrally disaggregating LANDSAT data. This involves isolating wavelengths that sugarcane most strongly reflects: green (B3, B4), near-infrared (B5), and short-wave infrared (B7).
-
Training the classifier. Using the processed 2015 data, a signature file is generated that distinguishes the spectral profiles of the landcover types provided.
-
Applying classification. I first do this with the 2015 data, but it can be applied longitudinally because the spectral profiles of the landcover types remain constant throughout time.
Finally, zonal statistics are extracted for the whole study region and exported to CSV files in google drive. With a detailed shapefile, this would yield data on, for example, the percentage of a district cultivating sugarcane for every year since 1984. This can easily be converted into km² or hectares with information about the area of each polygon.
There are two auxiliary files: AlgorithmComparison.js uses the same data to compare the performance of Google Earth Engine's different classification algorithms using a confusion matrix and F1 score. TrainingPolygons.js loads the training data used therein.