Investigating three-dimensional convolutional and recurrent neural networks for crop classification using time-series optical images

One of the serious challenges in remote sensing is extracting suitable features. With the presence of a new generation of deep neural networks, automatic and accurate feature extraction and classification of crops have become possible. On the other hand, appropriate features can partially reduce the effects of spectral similarity in the detection of different crops while improving classification accuracy. Also, the use of time-series data during the crop growth period provides useful information about crops to researchers. In this regard, this research aimed to investigate and evaluate three methods, three-dimensional convolutional neural network (3D-CNN), long short-term memory (LSTM), and gated recurrent unit (GRU), to extract appropriate features from time-series optical images. In the architecture of 3D-CNN, an attempt was made to design a structure so that the optimal spatial-temporal features could be extracted from time series images, and then the results were evaluated and compared with two other methods (i.e. LSTM, GRU). Finally, according to the results, 3D-CNN, with an overall accuracy (OA) of 90.70% and a kappa coefficient (KC) of 89.37%, which were about 3.50% and 4.00% higher than the OA and KC of LSTM, respectively, demonstrated a greater capability to identify crops. Moreover, the OA of the results of the classification by GRU was close to the OA of 3D-CNN, and only the OA of this method was 1.48% better than GRU. Therefore, the results confirmed the efficiency and suitability of 3D-CNN for crop classification.