Crop classification using aerial images by analyzing an ensemble of DCNNs under multi-filter & multi-scale framework

• Published in: Multimedia tools and applications 2023-05, Vol.82 (12), p.18409-18433
• Main Authors: Kalita, Indrajit, Singh, Gyan Prakash, Roy, Moumita
• Format: Article
• Language: English
• Subjects: Art techniques; Artificial neural networks; Computer Communication Networks; Computer Science; Data Structures and Information Theory; Datasets; Electronic devices; Feature extraction; Feature maps; Image classification; Image processing; Multimedia; Multimedia Information Systems; Neural networks; Special Purpose and Application-Based Systems
• ISSN: 1380-7501; 1573-7721
• DOI: 10.1007/s11042-022-13946-1

The rapid collection and development of multimedia data and devices allows agricultural monitoring to be automated. Crop classification using aerial images is the challenge of identifying the appropriate crop type planted on a certain area of land. These approaches, however, have several drawbacks that may inhibit the performance owing to a lack of appropriate data and the lack sufficient investigation of automatic feature extraction techniques. To deal with this, an ensemble approach has been investigated using two different deep convolutional neural networks (DCNN), namely Multi-Filter Multi-Scale deep convolutional neural network (MFMS-DCNN) and a pre-trained Inception V3 architecture, for crop classification using aerial images. At the onset, the MFMS-DCNN is constructed using Convolution- batchnormalization-activation units along with max and average pooling layers. Here, input images have been down-sampled using the max- pooling operation instead of traditional image processing approaches which facilitates to extract the important features in multiple levels, and the feature maps are concatenated by embedding a step-by-step fusion approach. The pre-trained Inception V3 model is initially fine-tuned using the three datasets under consideration (publicly available Plant seedling, and two new aerial image datasets obtained from two regions across India.). Finally, two combiners (Mean rule and Product rule) are enforced to ensemble the outputs of MFMS-DCNN and pre-trained Inception V3, to give the final prediction. To validate the feasibility of the proposed method, experimentations have been performed on the considered datasets and with state-of-the-art techniques. The findings derived from these image datasets produce superior performance for the proposed schemes as compared to the state-of-the-art techniques. Apart from that, with average accuracy ≈9% to ≈99% for all datasets, the ensemble methods with both combiners are proven to be more efficient than the two individual proposed schemes (MFMS-DCNN and fine-tuned Inception V3).