OBJECT BASED POSTPROCESSING METHOD FOR CROP CLASSIFICATION MAPS ACCORDING TO EACH CLASS SPECIFICITY

Obtaining reliable and accurate crop classification and land cover map based on satellite data, in particular high resolution data, is one of the most important tasks in remote sensing. Such maps provide basic information for many other applied problems and are vital in remote sensing studies. Despite of which machine learning methods were utilized for maps obtaining: traditional (Random Forest, Support Vector Machine, Multi-layer perceptron, logistic regression) or state-of-the-art approaches (autoencoder, convolutional neural network, recurrent neural network) there is some noise (single pixels or groups and clusters of pixels that wrong classify) on such maps. There are traditional methods for noise reduction, however these methods do not take into account image semantics. Therefore, they are not effective for filtration land cover and crop classification maps based on satellite images. The most complicated task in the filtering such maps is to preserve boundaries between different agricultural fields and to remove quite big clusters of incorrect classified pixels (objects) and at the same time save small farmer fields that are right classified. Thus, in this paper we proposed new method for postprocessing crop classification map based on algorithm that takes into account each class specificity and as a result utilizes different thresholds for different classes. We proposed investigate each object in classification map independently and decision should be: “is this whole object a noise or not?”. We consider each class independent and use connected component labeling technique for discriminating objects from classification map. Further different types of conditions based on sharpness and compactness where proposed for the investigated object. Accuracy and efficiency of this method with the proposed filtration method have been tested on the independent set and using the visual comparison with the results of utilizing common filters. Also, McNemar statistical test has been conducted to prove the statistically significant gain of utilizing proposed filtration methodology compare to common voting filte