Addressing computation resource exhaustion associated with deep learning training of three-dimensional hyperspectral images using multiclass weed classification

Addressing the computational bottleneck of training deep learning models on high-resolution, three-dimensional images, this study introduces an optimized approach, combining distributed learning (parallelism), image resolution, and data augmentation. We propose analysis methodologies that help train deep learning (DL) models on proximal hyperspectral images, demonstrating superior performance in eight-class crop (canola, field pea, sugarbeet and flax) and weed (redroot pigweed, resistant kochia, waterhemp and ragweed) classification. Utilizing state-of-the-art model architectures (ResNet-50, VGG-16, DenseNet, EfficientNet) in comparison with ResNet-50 inspired Hyper-Residual Convolutional Neural Network model. Our findings reveal that an image resolution of 100x100x54 maximizes accuracy while maintaining computational efficiency, surpassing the performance of 150x150x54 and 50x50x54 resolution images. By employing data parallelism, we overcome system memory limitations and achieve exceptional classification results, with test accuracies and F1-scores reaching 0.96 and 0.97, respectively. This research highlights the potential of residual-based networks for analyzing hyperspectral images. It offers valuable insights into optimizing deep learning models in resource-constrained environments. The research presents detailed training pipelines for deep learning models that utilize large (> 4k) hyperspectral training samples, including background and without any data preprocessing. This approach enables the training of deep learning models directly on raw hyperspectral data.