Crop type classification with hyperspectral images using deep learning : a transfer learning approach

Crop classification plays a vital role in felicitating agriculture statistics to the state and national government in decision-making. In recent years, due to advancements in remote sensing, high-resolution hyperspectral images (HSIs) are available for land cover classification. HSIs can classify the different crop categories precisely due to their narrow and continuous spectral band reflection. With improvements in computing power and evolution in deep learning technology, Deep learning is rapidly being used for HSIs classification. However, to train deep neural networks, many labeled samples are needed. The labeling of HSIs is time-consuming and costly. A transfer learning approach is used in many applications where a labeled dataset is challenging. This paper opts for the heterogeneous transfer learning models on benchmark HSIs datasets to discuss the performance accuracy of well-defined deep learning models-VGG16, VGG19, ResNet, and DenseNet for crop classification. Also, it discusses the performance accuracy of customized 2-dimensional Convolutional neural network (2DCNN) and 3-dimensional Convolutional neural network (3DCNN) deep learning models using homogeneous transfer learning models on benchmark HSIs datasets for crop classification. The results show that although HSIs datasets contain few samples, the transfer learning models perform better with limited labeled samples. The results achieved 99% of accuracy for the Indian Pines and Pavia University dataset with 15% of labeled training samples with heterogeneous transfer learning. As per the overall accuracy, homogeneous transfer learning with 2DCNN and 3DCNN models pre-trained on the Indian Pines dataset and adjusted on the Salinas scene dataset performs far better than heterogeneous transfer learning.