Madagascar - Antsirabé

Specific Project Objectives & Deliverables


The methodology for crop identification and crop area estimation is based on the combined use of object based image analysis and data mining. It involves 3 steps:

  1. Preprocessing steps: (i) All images were transformed to top of atmosphere reflectance (top of canopy reflectance was tested but without satisfactory results because of the absence of reliable information on atmospheric composition), (ii) the boundaries of each field/object of the field database were digitized based on very high resolution PLEAIDES imagery, (iii) SPOT DEM was processed to extract slopes.
  2. Building a learning database:  a learning database was built by extracting a set of 341 variables for each field/object of the field database including: 286 radiometric variables (spectral response, indices), 50 textural variables (only from PLEIADES imagery), and some topographic (altitude, slope) and geometric (object size) variables.
  3. Random Forest: (i) a classifier was built based on the learning database, for each level of the JECAM nomenclature, (ii) an optimization phase was performed by analyzing the importance (informative degree) of each variable used and reducing the amount of variables used to perform the classifications to an optimal volume (providing the best classification performances for each level), (iii) the classifications were performed for the whole site at each level.


Following figures present the class and overall accuracies obtained for each level of the JECAM nomenclature, and the maps obtained for the whole area at each level.


Table 1   Class and Overall Accuracies Obtained for each Level of the JECAM Nomenclature using Random Forest over the Learning Database

Figure 1   Crop-Non Crop Level

Figure 2   Land Cover Level

Figure 3   Crop Group Level

Figure 4   Sub-Class Level

Estimation of rice crop production

Work on estimation of rice crop production is in progress. Due to the small size of cultivated fields, compared to the spatial resolution of satellite images (10 – 20 m), temporal signal (NDVI from TOA reflectance in Red and PIR bands) extracted for each sampled plot was analyzed and smoothed using Stavitsky-Golay algorithm to eliminate noise linked to mixed pixels, but also to clouds, different sensors, and atmospheric effects. This allowed isolating only plots having a pure temporal signal (by comparing raw temporal signal and smoothed one). A set of twelve satellite variables was then extracted from the NDVI temporal profile of each plot: maximum NDVI of the growing season and integrals of the NDVI on different periods of the growing cycle. These satellite variables were compared to total biomass, dry biomass, full and empty grain yield to analyze the correlations. Results showed that the more the sorting was drastic (elimination of plots having a temporal NDVI profile with too much noise), the more the correlations between satellite and yield variables were good (but with less population). With a sorting leading to a resulting population of 14 irrigated rice plots (over the 88 plots initially available), the best correlations were obtained with the use of the integral of NDVI from the middle of the growing slope to the maximum of NDVI of the growing cycle (here referred to as integral.mid.max). Good and very significant (p ≤ 0.01) linear correlations were obtained between this integral.mid.max and total biomass (R² = 0.68) and straw biomass (R² = 0.61). For grain yield, the correlation was less important (R² = 0.58) but significant (p ≤ 0.01). Empty grain yield showed non-significant correlations with all satellite variables.

The estimation of rice crop production still needs to be analyzed further.

©2015 Joint Experiment for Crop Assessment and Monitoring