Evaluation of the Performance of Sar and Sar-Optical Fused Dataset for Crop Discrimination

Abstract

<p><strong>Abstract.</strong> Crop discrimination and acreage play a vital role in interpreting the cropping pattern, statistics of the produce and market value of each product. Sultan Battery is an area where a large amount of irrigated and rainfed paddy crops are grown along with Rubber, Arecanut and Coconut. In addition, the northern region of Sultan Battery is covered with evergreen and deciduous forest. In this study, the main objective is to evaluate the performance of optical and Synthetic Aperture Radar (SAR)-optical hybrid fusion imageries for crop discrimination in Sultan Bathery Taluk of Wayanad district in Kerala. Seven land use classes such as paddy, rubber, coconut, deciduous forest, evergreen forest, water bodies and others land use (e.g., built-up, barren etc.) were selected based on literature review and local land use classification policy. Both Sentinel-2A (optical) and sentinel-1A (SAR) satellite imageries of 2017 for Kharif season were used for classification using three machine learning classifiers such as Support Vector Machine (SVM), Random Forest (RF) and Classification and Regression Trees (CART). Further, the performance of these techniques was also compared in order to select the best classifier. In addition, spectral indices and textural matrices (NDVI, GLCM) were extracted from the image and best features were selected using the sequential feature selection approach. Thus, 10-fold cross-validation was employed for parameter tuning of such classifiers to select best hyperparameters to improve the classification accuracy. Finally, best features, best hyperparameters were used for final classification and accuracy assessment. The results show that SVM outperforms the RF and CART and similarly, Optical+SAR datasets outperforms the optical and SAR satellite imageries. This study is very supportive for the earth observation scientists to support promising guideline to the agricultural scientist, policy-makers and local government for sustainable agriculture practice.</p>

Reliable Crops Classification Using Limited Number of Sentinel-2 and Sentinel-1 Images

“…Accuracy of validation does not determine the actual accuracy of the product, which is the classification result. The phenomenon of accuracy overestimation using only validation data set, emphasized in the literature review [25], is confirmed in other literature, e.g., [7,20], and also in our research.…”

“…The overall accuracy of the single image classification was equal 79%, which is slightly higher then the value from the literature (77.4% [7], 77.2% [8]) and much better than 47% [16], the highest accuracy we obtained in May, a few weeks before harvest confirming [8], 6. Adding radar images did not improve the classification result, which is also confirmed in the literature [20,23], but due to the use of only one Sentinel-1 image, it does not allow us to generalize this conclusion and requires further research.…”

“…In most methods, it is not possible to obtain accuracy on the training set, but only on the validation and test set. In many publications the accuracy of validation (OA) is reported, which in almost all cases is above 80% (e.g., SV M = 97.7% [7], SV M = 98.96% [20], SV Mmodi f ied = 98.07% [13], RF = 93% [22], RF = 86.98% [23], RF = 83.96% [39], Dynamic Time Warping algorithm, NDVI time series classification = 72-89%, multi-band classification = 76-88% [40]). In some cases, the accuracy for test data is also delivered: 84.2% [7], 88.94% [20], which means in the case of 13.5% [7] less value than the accuracy of the validation and in the case of 10.02% [20] lower.…”

“…It is difficult to compare our experiment with the research design mentioned above. Note the number of training and test samples: 2005/341 points [7] and 2281/1239 pixels [20].…”

“…At the same time, the validation accuracy is determined on the basis of the samples from training set not used for learning. Some authors report validation accuracy and accuracy calculated on independent test data [7,12,20]. Others only provide information on accuracy based on external reference data not included in the training set [8,15,16,21].…”