A deep learning approach for early detection of drought stress in maize using proximal scale digital images

Neural computing methods pose an edge over conventional methods for drought stress identification because of their ease of implementation, accuracy, non-invasive approach, cost-effectiveness, and ability to predict in real time. To ensure proper irrigation scheduling and prevent major yield losses, the objective was to develop a deep learning (DL)-based custom convolutional neural network (CNN) framework for in situ identification and classification of drought stress in maize crops. An original image dataset was created by acquiring 2703 RGB images of maize crops under natural daylight conditions to incorporate noise and varied backgrounds. The dataset was augmented and divided in a ratio of 7:2:1 for the training, validation, and test sets. A custom-CNN model was built using feature blocks, fully connected layers, and dense layers, and compared with five state-of-the-art CNN architectures, i.e. InceptionV3, Xception, ResNet50, DenseNet121 and EfficientNetB1. The results revealed that the custom CNN model achieved accuracies of 98.71% and 98.53% on the training and test sets, respectively. In comparison, the ResNet50 and EfficientNetB1 transfer-learned CNN architectures achieved an equivalent accuracy of 99.26% each, followed by DenseNet121 with a 98.90% accuracy on the test set. The Xception model performed the worst, with the highest accuracy of 91.91% on the test set. The results demonstrate that the developed custom CNN model should be adopted for real-time implementation on resource-constrained edge devices because of the lower number of parameters (0.65 million parameters) compared to other state-of-the-art architectures.