SITS - Satellite Image Time Series Analysis for Earth Observation Data Cubes

Overview

The sits R package provides a set of tools for analysis, visualization and classification of satellite image time series. The main aim of SITS is to support land cover and land change classification of image data cubes using machine learning methods. The basic workflow in SITS is:

1. Create a data cube using image collections available in the cloud or in local machines.
2. Extract time series from the data cube which are used as training data.
3. Perform quality control and filtering on the samples.
4. Train a machine learning model using the extracted samples.
5. Classify the data cube using the trained model.
6. Post-process the classified images.
7. Evaluate the accuracy of the classification using best practices.

Installation

Pre-Requisites

The sits package relies on sf, terra and raster, which in turn, require the installation of the GDAL and PROJ libraries. Please follow the instructions for installing sf together with GDAL available at the RSpatial sf github repository.

Obtaining SITS

SITS is currently available on github, as follows:

# Please install the `sits` package from github
# and its dependencies
devtools::install_github("e-sensing/sits", dependencies = TRUE)
library(sits)
library(tibble)
library(magrittr)

AMI Image

For users that have an AWS account, we have prepared a set of AMI (Amazon Machine Images that are optimized for running SITS in the Amazon Elastic Compute Cloud (or EC2). The AMI has the following settings: SITS 0.9.6, Ubuntu 18.04, R 4.0.2, and Rstudio Server 1.3.959. All packages have been updated as of 21 August 2020. The AMI is available for the following regions:

• South America (sa-east-1)
• Frankfurt(eu-central-1)
• US East (us-east-1)
• Asia Pacific Singapore(ap-southeast-1)

When you create an EC2 instance based on this AMI, ensure that your ‘security group’ settings allow incoming HTTP (port 80), HTTPS (port 443) and SSH (port 20) traffic. After the EC2 instance is started, then copy-and-paste the ‘IPv4 Public IP’ address for your running instance to a web browser address bar. That should bring the RStudio server interface in your browser. Use “rstudio” as username and “e-sensing” as password.

Data Cubes

SITS has been designed to work with big satellite image data sets organised as data cubes. Data cubes can be available in the cloud or in a local machine. Currently, SITS supports data cubes available in the following cloud services:

1. Sentinel-2/2A level 2A images in AWS.
2. Collections of Sentinel, Landsat and CBERS images in the Brazil Data Cube (BDC).
3. Sentinel-2/2A collections available in Digital Earth Africa.
4. Data cubes produced by the “gdalcubes” package.
5. Local image collections organized as raster stacks.

SITS relies on STAC services provided by these cloud services. The user can define a data cube by selecting a collection in a cloud service and then defining a space-time extent. For example, the following code will define a data cube of Sentinel-2/2A images using AWS. Users need to provide AWS credentials using environment variables.

s2_cube <- sits_cube(source = "AWS",
                     name = "T20LKP_2018_2019",
                     collection = "sentinel-s2-l2a",
                     tiles = c("20LKP"),
                     start_date = as.Date("2018-07-18"),
                     end_date = as.Date("2018-07-23"),
                     s2_resolution = 20
)

In the above example, the user has selected the “Sentinel-2 Level 2” collection in the AWS cloud services. The geographical area of the data cube is defined by the tile “20LKP”, and the temporal extent by a start and end date. Access to other cloud services works in similar ways.

Users can derive data cubes from ARD data which have pre-defined temporal resolutions. For example, a user may want to define the best Sentinel-2 pixel in a one month period, as shown below. This can be done in SITS by the sits_regularize which use the https://github.com/appelmar/gdalcubes package. For details in gdalcubes, please see Reference [4].

gc_cube <- sits_regularize(cube   = s2_cube,
                           name          = "T20LKP_2018_2019_1M",
                           dir_images    = tempdir(),
                           period        = "P1M",
                           agg_method    = "median",
                           resampling    = "bilinear",
                           cloud_mask    = TRUE)

Accessing time series in data cubes

SITS has been designed to use satellite image time series to derive machine learning models. After the data cube has been created, time series can be retreived individually or by using CSV or SHP files, as in the following example.

library(sits)
#> SITS - satellite image time series analysis.
#> Loaded sits v0.11.0.
#>         See ?sits for help, citation("sits") for use in publication.
#>         See demo(package = "sits") for examples.
#> Using configuration file: /Users/gilbertocamara/Library/R/4.0/library/sits/extdata/config.yml
#> Additional configurations found in /Users/gilbertocamara/Library/R/4.0/library/sits/extdata/config_user_example.yml
# create a cube from a local file 
data_dir <- system.file("extdata/raster/mod13q1", package = "sits")
raster_cube <- sits_cube(
        source = "LOCAL",
        name = "sinop-2014",
        satellite = "TERRA",
        sensor = "MODIS",
        data_dir = data_dir,
        delim = "_",
        parse_info = c("X1", "X2", "band", "date")
)
# obtain a set of locations defined by a CSV file
csv_raster_file <- system.file("extdata/samples/samples_sinop_crop.csv",
                               package = "sits"
)
# retrieve the points from the data cube
points <- sits_get_data(raster_cube, file = csv_raster_file)
#> All points have been retrieved
# show the points
points[1:3,]
#> # A tibble: 3 x 7
#>   longitude latitude start_date end_date   label   cube      time_series        
#>       <dbl>    <dbl> <date>     <date>     <chr>   <chr>     <list>             
#> 1     -55.7    -11.8 2013-09-14 2014-08-29 Pasture sinop-20… <tibble[,3] [23 × …
#> 2     -55.6    -11.8 2013-09-14 2014-08-29 Pasture sinop-20… <tibble[,3] [23 × …
#> 3     -55.7    -11.8 2013-09-14 2014-08-29 Forest  sinop-20… <tibble[,3] [23 × …

After a time series is imported, it is loaded in a tibble. The first six columns contain the metadata: spatial and temporal location, label assigned to the sample, and coverage from where the data has been extracted. The spatial location is given in longitude and latitude coordinates. The first sample has been labelled “Pasture”, at location (-55.65931, -11.76267), and is considered valid for the period (2013-09-14, 2014-08-29). To display the time series, use the plot() function.

plot(points[1,])

For a large number of samples, where the amount of individual plots would be substantial, the default visualisation combines all samples together in a single temporal interval.

# select the "ndvi" band
samples_ndvi <- sits_select(samples_modis_4bands, "NDVI")
# select only the samples with the cerrado label
samples_cerrado <- dplyr::filter(samples_ndvi, 
                  label == "Cerrado")
plot(samples_cerrado)

Clustering for sample quality control

Clustering methods in SITS improve the quality of the samples and to remove those that might have been wrongly labeled or that have low discriminatory power. Good samples lead to good classification maps. sits provides support for sample quality control using Self-organizing Maps (SOM).

The process of clustering with SOM is done by sits_som_map(), which creates a self-organizing map and assesses the quality of the samples. This function uses the “kohonen” R package to compute a SOM grid (see Reference [7] below). Each sample is assigned to a neuron, and neurons are placed in the grid based on similarity. The second step is the quality assessment. Each neuron will be associated with a discrete probability distribution. Homogeneous neurons (those with a single class) are assumed to be composed of good quality samples. Heterogeneous neurons (those with two or more classes with significant probability) are likely to contain noisy samples. See Chapter 4 of the sits book.

Filtering

Satellite image time series are contaminated by atmospheric influence and directional effects. To make the best use of available satellite data archives, methods for satellite image time series analysis need to deal with data sets that are noisy and non-homogeneous. For data filtering, sits supports Savitzky–Golay (sits_sgolay()), Whittaker (sits_whittaker()), and envelope (sits_envelope()). As an example, we show how to apply the Whittaker smoother to a 16-year NDVI time series. For more details, please see the vignette “Satellite Image Time Series Filtering with SITS”

# apply Whitaker filter to a time series sample for the NDVI band from 2000 to 2016
# merge with the original data
# plot the original and the modified series
point_ndvi <- sits_select(point_mt_6bands, bands = "NDVI")

point_ndvi %>% 
    sits_filter(sits_whittaker(lambda = 10)) %>% 
    sits_merge(point_ndvi) %>% 
    plot()

Time Series classification using machine learning

SITS provides support for the classification of both individual time series as well as data cubes. The following machine learning methods are available in SITS:

• Linear discriminant analysis (sits_lda)
• Quadratic discriminant analysis (sits_qda)
• Multinomial logit and its variants ‘lasso’ and ‘ridge’ (sits_mlr)
• Support vector machines (sits_svm)
• Random forests (sits_rfor)
• Extreme gradient boosting (sits_xgboost)
• Deep learning (DL) using multi-layer perceptrons (sits_deeplearning)
• DL using Deep Residual Networks (sits_ResNet) (see reference [5])
• DL combining 1D convolution neural networks and multi-layer perceptrons (sits_TempCNN) (See reference [6])

The following example illustrate how to train a dataset and classify an individual time series. First we use the sits_train function with two parameters: the training dataset (described above) and the chosen machine learning model (in this case, extreme gradient boosting). The trained model is then used to classify a time series from Mato Grosso Brazilian state, using sits_classify. The results can be shown in text format using the function sits_show_prediction or graphically using plot.

# training data set
data("samples_modis_4bands")
# point to be classified
data("point_mt_6bands")
# Select the NDVI and EVI bands 
# Filter the band to reduce noise
# Train a deep learning model
tempCNN_model <- samples_modis_4bands %>% 
    sits_select(bands = c("NDVI", "EVI")) %>% 
    sits_whittaker(bands_suffix = "") %>% 
    sits_train(ml_method = sits_TempCNN(verbose = FALSE)) 
# Select NDVI and EVI bands of the  point to be classified
# Filter the point 
# Classify using TempCNN model
# Plot the result
point_mt_6bands %>% 
  sits_select(bands = c("ndvi", "evi")) %>% 
  sits_whittaker(bands_suffix = "") %>% 
  sits_classify(tempCNN_model) %>% 
  plot()

The following example shows how to classify a data cube organised as a set of raster images. The result can also be visualised interactively using sits_view().

# Retrieve the set of samples for the Mato Grosso region 
# Select the data for classification
# Reduce noise the bands using whittaker filter
# Build an extreme gradient boosting model
xgb_model <- samples_modis_4bands %>% 
    sits_select(bands = c("NDVI", "EVI")) %>% 
    sits_whittaker(bands_suffix = "") %>% 
    sits_train(ml_method = sits_xgboost())

# Create a data cube to be classified
# Cube is composed of MOD13Q1 images from the Sinop region in Mato Grosso (Brazil)
data_dir <- system.file("extdata/raster/mod13q1", package = "sits")
sinop <- sits_cube(
    source = "LOCAL",
    name = "sinop-2014",
    satellite = "TERRA",
    sensor = "MODIS",
    data_dir = data_dir,
    delim = "_",
    parse_info = c("X1", "X2", "band", "date")
)
# Classify the raster cube, generating a probability file
# Filter the pixels in the cube to remove noise
probs_cube <- sits_classify(sinop, 
                            ml_model = xgb_model, 
                            filter_fn = sits_whittaker())
# apply a bayesian smoothing to remove outliers
bayes_cube <- sits_smooth(probs_cube)
# generate a thematic map
label_cube <- sits_label_classification(bayes_cube)
# plot the the labelled cube
plot(label_cube, title = "Labelled image")

Additional information

For more information, please see the on-line book “SITS: Data analysis and machine learning for data cubes using satellite image timeseries”.

References

Selection of papers that use sits

• [1] Rolf Simoes, Michelle Picoli, et al., “Land use and cover maps for Mato Grosso State in Brazil from 2001 to 2017”. Sci Data 7, 34 (2020).

• [2] Michelle Picoli, Gilberto Camara, et al., “Big Earth Observation Time Series Analysis for Monitoring Brazilian Agriculture”. ISPRS Journal of Photogrammetry and Remote Sensing, 2018. DOI: 10.1016/j.isprsjprs.2018.08.007

• [3] Karine Ferreira, Gilberto Queiroz et al., Earth Observation Data Cubes for Brazil: Requirements, Methodology and Products. Remote Sens. 2020, 12, 4033.

Papers that describe software used in sits

We thank the authors of these papers for making their code available to be used in sits.

• [4] Appel, Marius, and Edzer Pebesma, “On-Demand Processing of Data Cubes from Satellite Image Collections with the Gdalcubes Library.” Data 4 (3): 1–16, 2020.

• [5] Hassan Fawaz, Germain Forestier, Jonathan Weber, Lhassane Idoumghar, and Pierre-Alain Muller, “Deep learning for time series classification: a review”. Data Mining and Knowledge Discovery, 33(4): 917–963, 2019.

• [6] Pelletier, Charlotte, Geoffrey I. Webb, and Francois Petitjean. “Temporal Convolutional Neural Network for the Classification of Satellite Image Time Series.” Remote Sensing 11 (5), 2019.

• [7] Wehrens, Ron and Kruisselbrink, Johannes. “Flexible Self-Organising Maps in kohonen 3.0”. Journal of Statistical Software, 87, 7 (2018).

R packages used in sits

The authors acknowledge the contributions of Marius Appel, Tim Appelhans, Henrik Bengtsson, Matt Dowle, Robert Hijmans, Edzer Pebesma, and Ron Wehrens, respectively chief developers of the packages “gdalcubes”, “mapview”, “future”, “data.table”, “terra/raster”, “sf”/“stars”, and “kohonen”. The code in “sits” is also much indebted to the work of the RStudio team, including the “tidyverse” and the “furrr” and “keras” packages. We also thank Charlotte Pelletier and Hassan Fawaz for sharing the python code that has been reused for the “TempCNN” and “ResNet” machine learning models.

How to contribute

The SITS project is released with a Contributor Code of Conduct. By contributing to this project, you agree to abide by its terms.