| Abstract |
In today’s digital age, the Internet of Things (IoT) is gaining popularity. IoT has made everything smart, for example, smart cities (smart buildings and smart homes), smart healthcare (personal monitoring, smart wearables), industrial automation (especially manufacturing), commercial (shopping systems, retail), and even that agriculture too. IoT is becoming increasingly important in today’s digital age, resulting in a rapid rise in the number of devices connected to it. Massive amounts of data will be produced by such widely distributed IoT devices at the network’s edge. Processing these massive amounts of data in the centralized cloud is expected to result in increased 400bandwidth utilization, latency, and network congestion. Edge computing has become a popular paradigm in recent years for reducing network congestion and serving real-time IoT applications by providing services close to enduser devices. Agriculture is the foundation of the economy of any nation in the world. By 2050, the world population will need a 70% increase in food production to feed an estimated global population of more than 9 billion people. Potatoes are consumed all over the world and its production plays an important role in agriculture. The two primary diseases that adversely affect the yield of potato crop production are early blight and late blight. Therefore, in this work, we used containerized microservices to deploy machine learning models to resource restricted edge nodes on agricultural land for real-time disease and irrigation water requirement prediction in potato crops. Containers are lightweight and easy to deploy, making them the ideal choice for running machine learning models on resource-constrained edge nodes. We examined AlexNet, MobileNet, and VGG16, three deep convolutional neural networks (CNN), to detect these diseases automatically. A dataset of 7128 images containing healthy and diseased leaves of potato plants was used to train all three CNN models on the cloud. Even if training is outsourced, trained models need a lot of RAM; hence, the first aim of this study is to find a lightweight CNN model that can easily fit into resource-constrained devices. To improve potato crop yield and reduce economic losses, we found and deployed lightweight CNN model at the edge node to identify diseases in real time using leaf pictures recorded by in-place camera devices. The advantage of the developed technique is that by classifying potato leaf pictures in real time on-premises, there is no need to transfer images to the cloud for probable disease identification, which increases network congestion. The lightweight CNN model achieved 99.87% accuracy for both train and test images, according to the results. Precision (P), Recall (R), and F1 score (F1) are also displayed, to visualize the model’s efficiency. Similarly for real-time prediction of irrigation water requirement in potato crop, we trained two machine learning models, Support Vector Machine (SVM) and Logistic Regression (LR) on 100,000 values of records. Each record has four input parameters (soil moisture, temperature, humidity, and how many days the crop was planted before). With four input parameters in each record (soil moisture, temperature, humidity, and how many days before the crop was planted), the model decides whether the water pump should be turned on or off. The SVM model and the LR model attained 92 and 73% accuracy, respectively, in determining whether the water pump should be turned ON or OFF. © 2024 by Apple Academic Press, Inc. |
