ESA Earth Home Missions Data Products Resources Applications
   
EO Data Access
How to Apply
How to Access
Services
Site Map
Frequently asked questions
Glossary
Credits
Terms of use
Contact us
Search


 
 
 

 

Supervised Crop Classification from Middle-resolution Multitemporal Images

Lukas Brodsky(1)

(1) GISAT, Charkovska 7, Praha 10, 101 00, Czech Republic

Abstract

The objective of this paper is to evaluate possibilities of land use - crop classification at pixel level applying artificial neural-network (ANN) technique using multitemporal Envisat Meris images. The main motivation is to provide classification of crops, or crop groups at pixel level to be used in crop growth monitoring systems or other models. One of the important aspects of the crop classification among others is extraction of phenological parameters from time series of vegetation indices. The availability of phenological data at regional level is crucial for e.g. model calibrations. There exists methodology on extraction of the parameters but it is important to know which pixel belongs to which crop or crop fraction area to derive correct parameters. This is very difficult task in case of 1 km images but use of middle resolution data (such as Meris with spatial resolution of 300 m) brings new perspectives. There have been initially selected three crop groups: winter, spring and summer crops. Additionally, winter rapeseed and other-crops classes were added. The last one was added to describe crops that exhibit higher variation, thus are hard to be defined and detected in 300 m pixels. This group includes mainly crops sowed in spring as vegetables, legumes and some of root and technical crops. The classification was done on area of the Czech Republic (about 77 000 km2) under mask of arable land (about 25 000 km2) coming from Land Cover map (30 m image based product). The agricultural regions cover mainly lowland intensive production areas at about 200 m above sea level to less extended middle land areas at about 500 m. Average parcel size is about 12 ha and maximum reaches the value of 300 ha. Pixels to be classified were selected with 75 % threshold of fraction area of arable land. Reference dataset (crop classification) for accuracy assessment consisted of ground survey and crop classification on multitemporal 20 m Spot images with overall accuracy 94.4 %. Classification models included linear, radial basis function (RBF) and multi-layer percepton (MLP) ANN with 50 networks tested in training. The training data set consisted of about 200 samples per class on which bootstrap resampling was applied. Selection of subset of independent variables (Meris spectral channels) was used in the procedure. The best selected ANN model (MLP: 3 in, 13 hidden, 3 out) resulted in very good performance (correct classification rate 0.974, error 0.103) applying three crop types data set. In the next step data set with five crop types was evaluated. The ANN model (MLP: 5 in, 12 hidden, 5 out) performance was also very good (correct classification rate 0.930, error 0.370). The first classification of the three crops compared with reference data set resulted in overall accuracy of 77.2 %. The extended classification of five crops reached overall accuracy of 68.1 %. By selecting reference data accounting > 75% fraction area of one crop in the pixel the overall classification accuracy increased by 6 %. The best classification accuracy was achieved for winter crops, 86.2 %. The purest accuracy was found in case of other-crops and spring crops. There existed the highest crops mixture resulting even in 61.6 % accuracy (spring crops). This may be caused by the fact that during the high vegetation period, when spring crops can be discriminated, the time series lacked good quality data given the atmospheric conditions (2007 season). Field structure under the 300 m pixels, hence mixed spectra, plays also considerable role. In the next step, neural net with mixture model will be tested for classification improvements. The study shows that classification of five crop types at pixel level of Meris data achieved about 70 % accuracy using ANN. Classification of these five crop types would be necessary to extract phenological parameters. Aggregating the results to regional administrative units, the accuracy is predicted to increase. For extraction of the parameters only pixels with high probability ratio can be selected and used.

Acknowledgements: The study was done under the framework of ESA CAT-1 (ref. 4358) and CGMS-CZ projects.

 

Workshop presentation

Keywords: ESA European Space Agency - Agence spatiale europeenne, observation de la terre, earth observation, satellite remote sensing, teledetection, geophysique, altimetrie, radar, chimique atmospherique, geophysics, altimetry, radar, atmospheric chemistry