Publications des agents du Cirad


Semiautomatic classification of sugarcane areas from Random Forest algorithm and OLI/Landsat-8 images

Luciano A.C.D.S., Picoli M.C.A., Rocha J.V., Silva A.C.D.O., Le Maire G.. 2016. In : Simposio Internacional en Percepción Remota y Sistemas de Información Geográfica. Puerto Iguazú : SELPER, 13 p.. Simposio Internacional SELPER 2016. 17, 2016-11-07/2016-11-11, Puerto Iguazú (Argentine).

Identification and mapping of sugarcane is prerequisite to mapping the crop production. It has been facilitated by several remote sensing data and classification techniques. In the present study, an object-based semiautomatic classification methodology was tested to classify sugarcane based on analysis of time series of 7 images (2013-2014 cultural cycle) from Landsat-8 OLI sensors. The study area is located in Piracicaba microregion in the Sao Paulo State, which is the principal producer state of sugarcane in Brazil. The Random Forest (RF) classifier was used to differentiate the sugarcane crop from other uses, like water, natural vegetation, pasture, urban areas, and other agricultural areas. Normalized Difference Vegetation Index (NDVI) and Enhanced Vegetation Index (EVI) were calculated from reflectance images of OLI sensors. These vegetation indices, together with the surface reflectance images from bands 2 to 7 (blue to far infrared light) were used as input variables of Random Forest classifier. Samples from representative areas of sugarcane and other uses, selected from maps done in the CANASAT project, were used to calibrate the RF classifier. After the calibration step, the classifier was applied in all temporal series of a segmented image. This segmentation was done under SPRING software and allowed to segment the image in more than 35x103 objects (polygones) of similar spectral aspect. The RF classifier attributes a single class for each of these polygons, thus resulting in a map of the available sugarcane areas to harvest from April. The sugarcane class in the study area, represented 32.1% of total area which corresponds to 253000 hectares of sugarcane available to harvest in April. Sugarcane map from CANASAT project was used to validate the results. The maps were compared through a confusion matrix and Kappa statistics. The results indicated that the semiautomatic classification achieved an overall accuracy of 83% and a Kappa coefficient of 0.61, which is reasonable taking into account that some uncertainty exist on CANASAT map. Random Forest classifier, applied to time series of OLI sensors images was adequate to identify sugarcane crop from other uses and land cover. However, additional analyses are necessary to better select the input variables and parameters of RF classifier to discriminate the crop.

Documents associés

Communication de congrès

Agents Cirad, auteurs de cette publication :