Deep learning opens up a plethora of opportunities for academia and industry to invent new techniques to come up with modified or enhanced versions of standardized neural networks so that the customized technique is suitable for any specialized situations where the problem is about learning a complex mapping from the input to the output space. One such situation lies in a farm with huge cultivation area, where examining each of the plant for any anomalies is highly complex that it is impractical, if not impossible, for humans. In this chapter, the authors propose an optimized deep learning architectural model, combining various techniques in neural networks for a real-world application of deep learning in computer vision in precision farming. More precisely, thousands of crops are examined automatically and classified as healthy or unhealthy. The highlight of this architecture is the strategic usage of spatial and temporal features selectively so as to reduce the inference time.
TopBackground
As evidenced by the detailed survey captured in (Radhakrishnan & Vijayarajan V., 2019), even as there are various researches happening in Deep Learning in the context of Farming 4.0, the economic viability of its application is still very less. The literature survey further provides research direction to optimize deep learning applications for monitoring crop health so as to make this technology useful to farmers. Hence the domain considered for this research work of optimizing deep learning system is “Precision Floriculture” in Green House cultivation where the produce is cut flowers.
That said, there are various researches being carried out to employ Deep Learning to classify plants based on their botanical parts like leaf, flower, etc. Han et al. (Han, Seng, Joseph, & Remagnino, 2017) investigated the use of deep learning to extract discriminatory features from images of leaves by learning and use them as classifiers to identify plant species. Their results demonstrate that compared to using hand crafted features of leaf images, learning the features using CNNs do provide better feature representations.
Gurnani et al. (Gurnani, n.d.) developed a deep learning framework to categorize flowers by doing transfer learning from two famous CNN architectures called GoogleNet which was pre-trained on ILSVRC2014 dataset and AlexNet which was pre-trained on ILSVRC2012 dataset and found their accuracies to be 47% and 43% respectively. By the way, ILSVRC stands for ImageNet Large Scale Visual Recognition Competition. In this study they confirmed the popular claim by researchers that the performance of the neural network will be commensurate with the depth of the network. This claim is further ascertained by Liu et al. (Liu, Yang, Cheng, & Song, 2018) who designed a very deep leaf classification model using a ten-layer CNN based on LeNet architecture to classify 32 species of plants and achieved an accuracy rate of 87.92%.
Further, the same LeNet architecture was used by Amara et al. (Amara, Bouaziz, & Algergawy, 2017) to build a deep neural network to detect two common banana diseases called banana sigatoka and banana speckle from banana leaves. Unlike other earlier experimental setups, they have used images captured under challenging conditions such as illumination, complex background, different images resolution, size, pose and orientation. The key take away from their work is that better accuracy of more than 90% is achievable using color images opposed to lower accuracy of about 85% using grey scale images. The authors achieved best performance with a train-test data split of 80%-20%.