isabella232 / satts

ML pipeline to classify crop types with multi-spectral and multi-temporal EO data

Pixel-level clustering and classification of multi-spectral, multi-temporal earth observation data

This library contains classes and functions to generate datasets corresponding to spatial features from a time-series of satellite images. The impetus for this project was to develop an easy to use, high-level interface to numerous Python modules for the clustering and classification of land cover/land use (LULC) types, with an initial focus on classifying individual crop types in challenging geographies using a time-series of multi-spectral earth observatoin (EO) images. The use of a time-series of EO images better captures the dynamic nature of the appearance of crops and other LULC classes through a growing season, enabling more accurate model predictions. The functions and methods provided in this library can be used to generate EO reflectance time-series datasets and models for arbitraty vector data, e.g. points or polygons.

Using this library

The library is divided in to several components:

1. tsmask: provides functions to create a masked numpy arrays corresponding to areas of interest, as well as a BandTimeSeries object initialized using the maked array. Specific functions and objects include:

   • raserize utilizes the osgeo library and the underlying gdal functionaility to rasterize vector features from a shapefile and output a .tif file sharing the relevant metadata and dimensions as the reference image from which it was created. A check_rasterize function is also provided to confirm that the features were correclty "buned" into the raster layer. The resulting image can be characterized as a land cover "mask".

   • mask_to_array generates a 3D numpy array from the output of rasterize. Each element of the 3D array is a 2D array representing band reflectance values for a given date. Values in the 3D array that are not no-data values correspond to a land cover class burned in using rasterize.

   • BandTimeSeries objects contain information about time-series' of reflectance values for samples in a given land cover class, and methods to operate on and format the reflectance time-series. BandTimeSeries objects are initialized using an output from the mask_to_array function, along with arguments specifying the land cover class of the object, and the variable (band) name of the reflectance time-series. The time_series_data_frame method allows for interpolation of the time-series.

2. tsclust: provides a TimeSeriesSample class that is useful for generating a dataset from all or a subset of data contained in a BandTimeSeries and formating it for direct use in the functions and classes provided in the tslearn library.

   • TimeSeriesSample take n_samples of the data in a BandTimeSeries and optionally smooth the time-series' using a Savgol signal smoothing. The ts_dataset method generates an object that can be used directly in the time series clustering and classification algorithms provided in the tslearn library.

   • cluster_time_series performs either GlobalAlignmentKernelKMeans or TimeSeriesKMeans (both from the tslearn library) on a TimeSeriesSample object. The user specifies the number of clusters as well as the distance metric used if the clustering algorithm is TimeSeriesKMeans (dynamic time warping or soft dynamic time warping). Sillhouette scores computed on the resulting clusters can optionally be returned. Alternative sets of hyperparamters for cluster_times_series can be tested using the cluster_grid_search function.

   • cluster_mean_quantiles and plot_clusters provide methods for inspecting and visualizing cluster results.

3. tstrain provides functions for extracting training datasets comprising time-series' of band reflectance values at known locations (x,y numpy array indices) from satelite scenes.

   • random_ts_samples takes n_samples from .csv files containging reflectance time-series data for a given land cover class.

   • get_training_data reads satellite scenes, e.g. scense corresponding to an areo of interest specified with sat-search and download and saved using the default direcorty structure ofsat-search load, into numpy arrays using functionaility from gippy. The output is a long-form pandas dataframe with colums for date, feature (band-value), band reflectance value, the 2d array index, and a label corresponding to a samples land cover class.

   • format_training_data takes the ouput of get_training_data and reshapes it into a 3D numpy array of shape (n_samples, n_timesteps, n_features) suitable for use in a Keras Sequential model. Both x and y (optionally one-hot encoded) are returned.

Examples

Coming soon: Two jupyter notebook tutorials showcasing the functionality in this library