Burkina Faso - Koumbia

Specific Project Objectives & Deliverables


Cropland/Crop Type identification:


(1) Build a cropland/crop type identification map at the highest possible spatial resolution (0.5m) provided by the available EO data for the 2014 agricultural season (data acquired in 2015 that will be processed in 2016), for the different levels of the JECAM nomenclature (see Table 1).

(2) Develop a novel methodology for classification leveraging the data-fusion approach and limiting the use of site-specific prior information, in order to devise a processing chain which can work at a global scale.

Methods: Our data-fusion approach relies on the OBIA (Object Based Image Analysis) paradigm:

  • an object layer is generated by segmenting the VHSR image, and a large set of radiometric/multi-temporal indices from HR imagery are projected at the object scale and joined to VHSR textural indices;
  • a Random Forest (RF) classifier is used to carry out object-based classification, as well as to perform importance analysis over selected variables and to assess classification performances;
  • several classification strategies have been tested to match the different levels of the JECAM nomenclature with an additional Level 0 for crop vs. non-crop identification (see Table 1), namely a traditional single level approach, and two (top-down and bottom-up) hierarchical approaches;
  • a more robust validation is also performed using external validation segments not included in the training set, obtained by photo-interpretation based on additional in-situ data.

Table 1   Modified JECAM Nomenclature including Level 0 (Crop vs Non Crop)

Variable Importance Analysis

To limit the complexity of the overall methodology, we performed a first set of Random Forest classifications to assess the number of important variables to retain. The figure and table below (Figure 1) show the overall accuracies as a function of the number of important variables used for levels 0, A and B. These experiments confirm that a number of variables around 20 (one tenth of the total) is enough to achieve a satisfactory accuracy (above 95% of the maximum achievable accuracy).

Figure 1   Overall Accuracies Obtained Varying the Number of Important Variables Used


Classification assessment using internal Random Forest validation

A first assessment of classification accuracies has been carried out using the internal Random Forest validation strategy (mean of the accuracies on randomly chosen validation samples over different trees). Encouraging results have been obtained, especially for the Levels 0 and A, as reported in Figure 2. Scores for the most detailed levels C and D are very promising, but further inspection is necessary to confirm these outcomes.


Classification assessment using external validation segments

A further set of manually segmented areas (mainly obtained by photo-interpretation) has also been used as an additional test set to assess classifications. Accuracies obtained using this test set are less interesting, especially starting from level B, as shown in Table 2. However, the reliability of the external validation set has to be further inspected.

We could also test the different classification strategies, and verify that the hierarchical approach starting from the level-0 map gives the best accuracies at finer scales.


Figure 2   Overall Accuracies for Single Level Classification using Internal RF Validation



Level 0

Level A

Level B


87.4 %

87.4 %

50.5 %



90.5 %

54.1 %

By grouping


84 %

37.1 %


Table 2   Overall Accuracies for Different Classification Strategies using External Validation Data

In the next Figures, some samples of the cropland/crop-type maps generated for the 2014 agricultural season are shown.

Figure 3   Level 0 Map (Crop vs Non Crop) for Koumbia Village

Figure 4   Details of Classifications at Levels 0, A and B

Yield Prediction and Forecasting:


(1) Describe and evaluate the main crop systems of the site: crop varieties, crop rotation, use of inputs, tillage, fallow, use of plough or tractors and

(2) quantify the yield variability obtained by farmers and evaluate the link with the climate variability.

Methods: Six villages have been selected according to their spatial distribution, their accessibility, the studies already carried out, and the remote sensing image footprints. In agreement with the farmers and peasant organizations, thirty plots have been chosen in each village, to carry out two types of survey:

  • A survey with the farmer, concerning the plot monitored: Preceding crops for the three last years, crop management techniques, area cultivated, production obtained, crop residues.
  • Concerning the crop monitoring on the plot for the season 2014:
    • Ten days of crop monitoring
    •  The weighing of grains and biomass for three quadrants by plot.
    • The daily measurement of rainfall, with three rain gauges put in each village (a total of eighteen rain gauges).

Figure 5   Site with Villages and Monitored Plots

Figure 6 shows the cotton yields for each village in recent years, and Figure 26 shows the crop rotation with maize.

Figure 6   Cotton Yields for each Village, 2011 - 2013

Figure 7 Crop Rotation with Maize 2009-2011

Figure 8 shows the rainfall at the rain gauges in Boni village, and Figure 9 shows the sowing related to rainfall in Gombeledougou.

Figure 8   Rainfall at the 3 rain gauges in Boni Village, 2014


Figure 9   Sowing Related to Recorded Rainfall in Gombeledougou, 2014

Data analysis for 2015 growing season is still ongoing. The number of monitored plots has been raised to 160, 85 cultivated with maize, 33 with cotton and 42 with sorghum.


An appropriate general workflow based on the fusion of heterogeneous data has been successfully carried out for the identification of crop types. Current classification scores, although to be further validated, stand unprecedented for the Burkina Faso site and confirm that the proposed approach is promising. However, further development has to be carried out in order to:

  • select more appropriate variables at different scales;
  • refine methodology with respect to object-layer generation and object-based classification strategies;
  • collect and process data from other sensors (radar) and/or sources,
  • identify a more reliable external validation strategy.

We followed the recommendations of the “JECAM guides” for the acquisition of field data. However, we have adapted to the nomenclature cultures present on our site.

We modified the project objectives in the sense that we added the study of Yield Prediction and Forecasting.

©2015 Joint Experiment for Crop Assessment and Monitoring