Lesson 1, Topic 3
In Progress

Data Collection and Processing v2

Lesson Progress
0% Complete

This topic deals covers ways of how SAR data and ground observations are integrated to better map Canada’s agricultural resources. Here, Thierry Fisette explains how AAFC collects, processes and integrates satellite imagery and field data to create Canada’s Annual Crop Inventory.

Input Data: Ground Truth Information

Ground reference, or in situ, information is critically important for training and validating models and successful crop classification. Currently, annual crop insurance data are provided by [[ or for?? .. need to check entire paragraph! ]] four Canadian Provinces – Alberta, Saskatchewan, Manitoba and Québec – representing 87 per cent of Canada’s agricultural area. The crop insurance data sets are the most accurate, detailed and complete sources of geospatial crop information in Canada.


For provinces where insurance data cannot be accessed, ground-truth information is provided by in-situ observations from AAFC staff. In-situ data are screened and evaluated for bias by before they are inputted into the DT classifications. In 2019, AAFC personnel gathered 70,981 observation points in Canada. 70% of these observations were used to train the classifier and the remaining 30% were used as a validation dataset on the final map product.

Here comes the real GIF!!

Field data collection

here is a video…..


By loading the video, you agree to YouTube’s privacy policy.
Learn more

Load video


By loading the video, you agree to YouTube’s privacy policy.
Learn more

Load video

Input Data: Optical & Radar Imagery

Optical multi-spectral data such as Landsat-8 and Sentinel-2 are adequate to classify crops, if data are available during critical periods of the growing season. Accuracies greater than 85% have been achieved (overall and for individual crops), but can be significantly degraded by gaps in optical data collection. Results from research and operations have shown that the integration of SAR can increase accuracies over the use of optical data alone. For example, the addition of the dual-polarization RADARSAT-2 data to optical imagery has increased overall accuracies by up to 16 percent. Both optical and SAR data streams provide unique and valuable information relating to plant growth and type. Optical imagery is important because observations acquired in the NIR and SWIR regions of the electromagnetic spectrum have been shown to be extremely useful for differentiating among crop types. Dual-polarization SAR data is much more sensitive to plant structure than optical data and is also able to fill gaps in the optical image record brought about by non-ideal weather conditions during key growth stages.

Before using optical and radar data as inputs in to the DT classification used to generate the ACI, some pre-processing of these data must be carried out. If not already geometrically corrected, all images are orthorectified using a 3-D multi-sensor physical model. Bottom of Atmosphere (BOA) reflectance images are derived from optical data and, for the 2020 crop inventory, compact polarimetry variables such as M-Chi decomposition, circular polarizations and Stokes vectors, will be derived from medium resolution (30m), 125 km swath width, RADARSAT Constellation Mission (RCM). More than 1500 RCM images or 1 complete national acquisition per month for 5 months, will be ingested into the classification process. A sigma Naught radiometric calibration and a gamma maximum-a-posteriori filter are applied on Dual-Pol (VV, VH) Radarsat-2 wide mode data to remove noise (speckle). For both radar and optical data, along-track images of the same date are mosaicked together. This process allows the creation of large classification regions with more training sites.

Decision Tree Classification

Since 2009, the AAFC has created its Annual Crop Inventory (ACI) by applying the see5 algorithm to various combinations of optical (Landsat-5, -7 and -8; Resourcesat-1; DMC; SPOT; Sentinel-2) and SAR (RADARSAT-2; RCM) imagery using discriminate functions estimated empirically from hundreds of thousands of ground-based (in situ) training data samples. The C5.0 Decision Tree (DT) method is preferred because of its advantages over other methods and, with particular reference to AAFC operations, its demonstrated ability to handle discrete data, its processing speed, its independence of the distribution of class signatures, its interpretable classification rules, its cost-effectiveness and demonstrated higher accuracies. AAFC has also incorporated advanced options, such as pruning and boosting, into the DT classification process to improve the accuracy of the algorithm.

AAFC performs its annual DT classifications on a region-by-region basis. This is because the dynamic nature of crop rotations, crop growth and harvest patterns create significant reflectance differences between adjacent satellite scenes within the temporal period encompassed by scene availability. While each classification region combines several dates of optical and SAR imagery, the actual combination of imagery per region can vary based on data availability.

Accuracy of Annual Crop Inventory

Although the ACI consistently achieves an overall target accuracy of 85 percent at the national scale, map accuracy varies from crop to crop, region to region, and year to year, depending on satellite data availability and the geographic representativeness of in-situ training data. In general, the highest mapping accuracies (greater than 90 percent) are found where crops display significantly different spectral characteristics at the time of the EO data acquisition, such as in the Canadian Prairie Provinces of Saskatchewan and Alberta. Elsewhere, however, accuracies are lower, and may vary from approximately 75 to 85 percent. This variability is explained by two main factors that relate to limitations associated with the in-situ and satellite data used in the classification process.

First, there can be significant Province-to-Province variability in the number, density and quality (i.e. detail and accuracy) of in-situ data used for training the Decision Tree (DT) classification and its validation. This variability is directly related to the different sources of the data.
The highest quality data tend to be those that:

  • are the mostly spatially representative of the agricultural landscape,
  • are of the highest accuracy through both space and time,
  • use the most detailed crop (thematic) classes; and
  • contain the large sample sizes required for training and validating the DT classifications. Data provided by the Provincial Governments of Alberta, Saskatchewan, Manitoba and Québec generally meet these all of these requirements. In comparison, data for the remaining Provinces (British Columbia, Ontario, New Brunswick, Nova Scotia, Prince Edward Island and Newfoundland) are collected by AAFC and provincial staff using windshield surveys that provide a less detailed and less spatially representative picture of the agricultural landscape as a whole.

Second, there is often significant regional variability in cloud cover, which can limit the amount of optical satellite data (currently Sentinel-2 and Landsat-8) available for input to the DT classification. For example, while as many as 3 or 4 clear-sky optical images can be acquired over the Canadian Prairie Provinces each growing season, this availability can be reduced to a single image (or, in the worst cases, no images) in Canada’s coastal regions (mainly the agricultural regions in the eastern Maritime Provinces of New Brunswick, Nova Scotia, Prince Edward Island and Newfoundland). While AAFC places a heavier reliance on the monthly acquisition of microwave (RCM and RADARSAT-2) imagery for input to the DT classification process in these regions, the lack or absence of optical imagery still results in a 10-15% drop in map accuracy compared to other regions where optical data is more available. This issue should become less problematic in the future as access to new streams of optical and microwave imagery increases.


Final Maps Distribution

As part of the Canadian federal government commitment to open data, the annual crop inventory datasets are publically available at no cost at http://open.canada.ca. The crop inventory mosaics are distributed in raster format (GeoTIFF) in the Albers Equal Area projection. The AAFC metadata – information that accompanies and describes the dataset – are based on ISO 19131 standards. Bilingual (French and English) metadata are available. To date, annual crop maps are typically delivered around six months following the end of growing season.


Further references and links related to this sections are located here.

Next Lesson:

Lake and River Ice

With this topic we end the lesson about mapping and monitoring of Canada’s agricultural landscapes. Move on to the next lesson to learn about how Canada monitor’s its lake and river ice with SAR sensors.

Please complete the quiz below to advance to the next lesson.

Click the ‘Submit’ Button to rate this topic.