Crop classification with deep convolutional neural network based on crop feature

Introduction:
Given that agriculture has the most important role in ensuring food security (Johnston & Kilby,1989), it is necessary to prepare a map that shows the spatial distribution, land area, and type of crops cultivated with high accuracy (Cai et al., 2018). Agricultural land cover is relatively dynamic and variable at relatively short intervals. This makes it difficult to classify crops on satellite imagery (Bargiel, 2017). The lack or absence of ground truth data is another cause. Therefore, methods that are less dependent on ground samples and use phenological features derived from time series of bands and vegetation indices to classify crops will be more appropriate (Ashourloo et al., 2020). The purpose of this study is to use a deep learning method based on convolutional networks to classify the crop types and improve the performance of this network by using feature channels as an input image to the network and increasing the classification accuracy.
Materials and methods:
In this study, the visible and near-infrared bands of Sentinel-2 satellite on 10 different dates from 2019 for an area in Idaho, USA, as an important agricultural area, and the cropland data layer for extracting the crop types ground labels was used (Han et al., 2012). Then, in MATLAB software, the time series of spectral bands were constructed and using them, temporal profiles of NDVI for any crop were extracted to identify the unique phenological features of crops. Then, the functions developed based on the phenological characteristics of crops were applied to the time series of the bands and a feature channel was obtained for each crop that in two separate processes, once bands and once again feature channels were used as input to the CNN and the network was trained and the results of network performance on crop classification in the test site, were compared.
Results and discussion:
In the first stage, the time series of bands formed the input of the deep convectional neural network and the network was trained in the training area, using the tempo-spectral information of bands as the input channels and crops ground samples as the related labels. Due to the spectral overlap of the crops in some time periods, network training was associated with a relatively high loss and therefore, for the test area, the overall classification accuracy was 69% (percent) and the kappa coefficient was 0.55. In the next step, the functions that were developed as phenological features for crops were applied on the time series of the bands, and for each crop, a feature channel was obtained as the special feature of that crop. Then the algorithm was implemented using these feature channels in the test area and the overall accuracy was upgraded to 86% and the kappa coefficient to 0.82 compared to which indicated a significant improvement in the results compared to the previous case.
Conclusion:
The deep convolutional neural network is very sensitive to the type of input channels for detecting agricultural crops and selecting the channels with suitable tempo-spectral characteristics for different types of crops, has a great impact on the accuracy of network training and can reduce the loss of training network and increase its efficiency in the classification of various crops.