Vegetable mapping using fuzzy classification of Dynamic Time Warping distances from time series of Sentinel-1A images

Abstract

Vegetable production is important because of the food security, diet improvement and socio-economic value. Mapping the location and extent of vegetable fields is therefore important in agricultural policy, food security and farmer support. Dynamic Time Warping (DTW) is a common way to map crops from time series of satellite images. However, as all hard classifications, it does not show the spatial distribution of uncertainty in the classification. In fuzzy classification, where memberships to multiple classes are assigned to each pixel, differences in membership between the first and the runners-up class can be used to assess classification uncertainty at the pixel level. This research formulates a fuzzy classifier based upon Time-Weighted Dynamic Time Warping (TWDTW) distances to map vegetable types from time series of Sentinel-1A SAR images. For each pixel, the TWDTW distances to the classes was normalised by dividing them by the sum of all TWDTW distances to all the classes for that pixel. The normalized distances were then used to compute fuzzy memberships for each pixel to each class, using the Gaussian membership function. Based on these memberships, fuzzy measures such as Confusion Index (CI), Ambiguity Index (AI), fuzziness and fuzzy membership were calculated and different thresholds applied on each of the measures during subsequent defuzzification. The overall accuracy and kappa coefficient of the defuzzified output results were 0.86 and 0.83, respectively, which was an improvement with regard to the crisp Time-Weighted Dynamic Time Warping with SPRING strategy for subsequence searching (TWDTWS) algorithm with 0.73 and 0.68 for overall accuracy and kappa, respectively. This study concludes that this new approach improves classification accuracy in image classification by excluding pixels with high uncertainty, which is especially relevant when only a limited number of classes are sampled and mapped.