LAUSR

ABSTRACT Crop mapping is a challenging task due to the spatial, spectral, and temporal variations within the cropland. These variations cause high intra-class and low inter-class variability problems. In this study, inspired by Deep Learning (DL) techniques, two Auto-Encoder (AE)-based learning schemes are proposed to exploit the spatio-temporal features in order to increase the stability of remotely sensed data classification for crop mapping. The first strategy is based on stacking the spatio-spectral features of different imaging dates to provide spatio-temporal spectral features and then feeding them as input to the Stacked AEs (SAEs). The spatio-spectral features are achieved by stacking the spectral features of all pixels in a Neighbourhood Window (NW) in each imaging date. The second strategy is an ensemble learning scheme, in which the base classifiers are SAEs trained considering different NW sizes. This method has an advantage such that while using the neighbourhood information, it maintains the classification accuracy in the boundary areas. To evaluate the performance of the proposed strategies, they were compared to the conventional classifiers, namely Linear Support Vector Machines (SVMs), Gaussian SVM, and Random Forest (RF). In addition, the effect of different train data sampling strategies and different proportions of train data on the classification performance was examined. The proposed ensemble (EN) strategy with a configuration of combining three NW sizes of 1, 3, and 5 pixels (i.e., ‘AE, EN (1, 3, 5)’) reached the highest accuracy in all experiments. Notably, in the experiment with a set of four imaging dates and 5.0% train data, it achieved an Overall Accuracy (OA) of 95.26%, the Kappa coefficient (K) of 0.94, and the average class accuracy of 92.16%.