Brazil - São Paulo

Specific Project Objectives & Deliverables

The method for mapping land cover in this site was described in the JECAM poster presented at Kiev:

We have seen from the dec 2014 results that it is possible to classify the perennial crops and forests with a good precision, with a single class “annual crop”. Once the “annual crop” mask generated in the first step, we will improve its sub-classification in a second step.  There is a big challenge for the classification of the cropland areas in the level of species, as we have seen last year: the lowest accuracies was for the annual crop species. The different options are: perform a classification with a reduced number of species, by grouping species types (eg. winter cereal). Another option is to classify the entire annual rotation, with a reduced number of rotation obtained from the table with a merge similar type of rotations. The last option is to perform a date-by-date classification at the species scale, however, some species have few training points. It was however the method that we selected, but grouping some species together. Indeed, the diversity of intra-annual crop cycles are very variable.

Another main question is which remote sensing data could be used. Indeed, during the wet season, very few images are available. Therefore, we are developing a new classification program which could use all the available data, including Landsat 7 data or images only partly covering the area. This is a challenge mainly for the issue of the “no data” values within the time series (clouds, area covered by the satellite, Landsat 7 SLC issue, etc.). For this, we will test two options: 1) filling the “no-data” with an advanced gap filling algorithm, based on the available data at that period and the surrounding data 2) training different classification models in function of the available data. The second algorithm showed better results and was kept for the rest of the study.

Annual land cover map :

The global accuracy for the 11 class nomenclature was 0.88 on the calibration dataset (see confusion matrix below) for February 2016. The highest confusion was between sugarcane and pasture, orange orchards and other perennial classes, and bare soil and annual crops. However, the confusion matrix is not independent from the calibration dataset, and accuracies may therefore be overestimated.

 

Visual interpretation of the obtained map show good prediction for most of the classes. The clear distinction between sugarcane area in the NW and E, forest plantation in the Center West, annual crops in the south and pastures in the Center East is largely coherent with other large scale maps of the area such as the Probio Map (http://www.mma.gov.br).

The orange orchards are well predicted for large fields, but there is a confusion with natural vegetation for small polygons.

Crop map at 3-month time step: The global accuracy was computed for each field inventory dates, considering only the crop class that had more than 10 observations in the calibration dataset (see tables below). The calibration dataset is therefore largely reduced to 115 to 222 observations. The global accuracies are given in the figure below, and highly depends on the date.

Conclusions

While the algorithm used was able to deal efficiently with clouds or other no-data, there are at least three main issues with this ongoing work :

  • The confusion in some classes (e.g. orange orchards) which can be solved either by adding field observation in the training dataset, or adding better predictive variables. The segmentation step can also have a high impact on the classification result, especially in small and fragmented areas such as riparian forests, which should be quantified
  • The crop types have been predicted for each date independently, which reduces the number of observations for training the algorithm, and reduced also the number of class
  • The “bare soil” class come with a high confusion with crop. Indeed, in first growth stages of crops the proportion of visible bare soil is high. In late growth stages, many crops become dry and are confounded with residues in harvested fields.

This JECAM site is particularly interesting to test classification algorithm including perennial tree-based agriculture (eucalyptus, pines, coffee, orchards), pastures, perennial crop (sugarcane), and annual crop.

©2015 Joint Experiment for Crop Assessment and Monitoring
© HER MAJESTY THE QUEEN IN RIGHT OF CANADA SA MAJESTE LA REINE DU CHEF DU CANADA (2015)