Abstract

CONTEXT

     Automated monitoring of the soil-plant-atmospheric continuum at a high spatiotemporal resolution is a key to transform the labor-intensive, experience-based decision making to an automatic, data-driven approach in agricultural production. Growers could make better management decisions by leveraging the real-time field data while researchers could utilize these data to answer key scientific questions. Traditionally, data collection in agricultural fields, which largely relies on human labor, can only generate limited numbers of data points with low resolution and accuracy. During the last two decades, crop monitoring has drastically evolved with the advancement of modern sensing technologies. Most importantly, the introduction of IoT (Internet of Things) into crop, soil, and microclimate sensing has transformed crop monitoring into a quantitative and data-driven work from a qualitative and experience-based task.

OBJECTIVE

     Ag-IoT systems enable a data pipeline for modern agriculture that includes data collection, transmission, storage, visualization, analysis, and decision-making. This review serves as a technical guide for Ag-IoT system design and development for crop, soil, and microclimate monitoring.

METHODS

     It highlighted Ag-IoT platforms presented in 115 academic publications between 2011 and 2021 worldwide. These publications were analyzed based on the types of sensors and actuators used, main control boards, types of farming, crops observed, communication technologies and protocols, power supplies, and energy storage used in Ag-IoT platforms.

RESULTS AND CONCLUSION

     The result showed that 33 variables measured by various sensors were demonstrated in these studies while 10 actuations were successfully integrated with Ag-IoT platforms. Perennial crops, which introduced less disturbance to Ag-IoT platforms than annual crops, were selected by 64% of researchers. Furthermore, studies in Ag-IoT system development were more focused on outdoor than indoor environments. Ag-IoT systems based on Arduino were most common among the studies while commercial platforms were least adopted, likely due to their inflexibility in customized developments. More researchers focused on agricultural applications than the IoT technology itself. Soil water content-based irrigation scheduling and controlled environment monitoring and controlling were the main applications. Other application areas included soil nutrient estimation, crop monitoring based on multiple vegetation indices, pest identification, and chemigation.

SIGNIFICANCE

     Several potential future research directions were identified at the end of the review, including integration of satellite-based internet connectivity to improve the IoT networks in non-connected farms, development of mobile IoT platforms (drones and autonomous ground vehicles) with continuous connectivity, and the use of edge-computing and machine-learning/deep-learning to enhance the capability of the Ag-IoT systems.

Introduction

     Crops are essential for human life because they provide food, animal feed, fuel, and raw materials for clothing and shelter. Crop yield has to be doubled in 2050 compared to 2009 in order to meet the demand of a growing population while increasing the food quality and reducing production inputs (Fukase and Martin, 2020). Potential solutions to enhance global food security include closing crop yield gaps, reducing food waste, changing dietary habits, and reducing inefficiencies in resource use (Foley et al., 2011). Reducing inefficiencies in input resources (such as water and nitrogen) can be achieved by continuously monitoring crops, soil, and microclimate, and then properly controlling inputs without sacrificing the yield and quality of the crop. Internet of Things (IoT) becomes a key technology that enables continuous monitoring and control in this scenario. The ability to generate (near) real-time quantitative data with high spatiotemporal resolution is a major advantage of IoT systems (Liao et al., 2017). IoTs are considered big data systems due to volumes, velocities, and varieties of data they generate. These data are mined and modeled to elucidate the relationships between inputs and outputs (Tsai et al., 2014). Correlation, trend analysis, classification, and numerical prediction are implemented on the data to reach meaningful control decisions. Compared with the conventional wireless sensor networks, the holistic approach of IoT technology allows users to incorporate data analytics on the big data collected by IoT sensor devices. Generally, connected actuators are enabled to control the inputs to achieve desired application rates. For example, an internet-connected soil water content (SWC) sensor network measures the plant water deficit and uploads data to a cloud-based data analysis platform. The analysis will find the trend of soil water deficit to determine the best time and quantity to apply irrigation water.

Conclusions and future directions

Conclusions

     Ag-IoT is a promising technology that would increase resource use efficiency in agricultural systems, and is an essential tool for digital agriculture transformation. In this paper, we have overviewed impactful research related to Ag-IoT in the past decade. The data collected from these papers were categorized and analyzed under six main Ag-IoT system design parameters namely sensors, sensing platforms and main control board, communication technology and IoT protocols, cloud platforms, power and energy management, and actuators. According to the analyzed data, it is revealed that there is an increased global attention towards the Ag-IoT system-related research in the recent years. However, there are certain research gaps found in the literature. One of them is that the implementations of the sensors and the actuators seem to be limited to soil and environmental parameter monitoring and irrigation controlling. Furthermore, crop macro and micronutrient demand analyses are still at the infant stage due to the non-availability of sensors that can measure nutrients in real-time. Therefore, it is essential to improve the sensor and actuator applications in crop monitoring and controlling. In addition, heterogeneity of the system parameters (such as data, platforms, required power) is a major challenge to the Ag-IoT systems implementation, to which the improvement of the context-awareness could be a solution. Power harness options for Ag-IoT nodes need more exploration as there are limited options available and it would be a big advantage for the perennial crop monitoring. The perception and the network layers of Ag-IoT systems require more improvements to meet the sensor implementation ergonomics and long-range high-throughput data transmission, respectively. Edge computing can be a replacement of the high throughput long-range communication, but to the best of our knowledge, only a limited number of practical applications have been developed based on edge computing to date. Mobile Ag-IoT platforms such as unmanned aerial and ground vehicles have a huge potential to increase the spatiotemporal resolution in Ag-IoT-based monitoring and controlling.

Abstract

     Recent advances in communication technologies with the emergence of connected objects have changed the agricultural area. In this new digital age, the development of artificial intelligence, particularly deep learning, has allowed for acceleration and improvement in the processing of collected data. To highlight the evolution and advances observed in deep learning in agriculture, we conducted a bibliometric study on more than 400 recent research studies. The analyses carried out on recent research works suggest that deep learning is widely involved in the digitization of agriculture areas with high accuracy exceeding the standard image processing techniques. Most of the works focus on crop classification problems, weed, and pest identification. Their methods are mainly based on convolutional neural network architecture. From the cases study, we have identified three key challenges that are essential in the deep learning methods applied in agriculture: (i) the need to consider the perception of the domain actors, their appropriation or interaction with the existing tools; (ii) the requirement to perform statistical tests to analyze the performance of the classifiers resulting from the learning process; and (iii) the need to perform statistical cross-validations with the training data. In the end, we summarized the agricultural data processing process consisting of several parts, for a better consideration of the expectations resulting from the challenges addressed. We consider that this study can serve as a guideline of research for the scientist and practician in the application of deep learning methodology in agriculture.

Introduction

     Agriculture is the practice of growing food. Given the increase in population, this sector must meet many requests for food while considering societal (such as labor), environmental (water scarcity, loss of biodiversity, land degradation, etc.), and economic issues. The constraints of its development have become numerous given the seasonal variability and extreme climate. It is more necessary than ever to find innovative practices for the development of the agricultural sector.

     Digital integration has significantly changed farmers’ knowledge of field management with innovative technologies like intelligent computing, robotics, drones, or sensors onboard farm machinery. Using these technologies is encouraging data scientists and agronomists to design analytical tools and techniques to accurately organize field management and address the new challenges at hand (fungal attack detection, crop yield prediction, advanced spraying, etc.). These novel practices require farmers’ technical assistance to support their needs and help them maximize their crop yields based on data and task automation.

Conclusion

     Precision agriculture is a recent science that is certainly still in its young phase. Deep learning can make a decisive contribution to the analysis of agricultural data, in this case by using computer vision to enable a machine to automatically analyze and understand the visual world. This offers the opportunity to develop intelligent systems, which appear to be one of the many possible ways to tackle the economic, environmental, and social challenges in agriculture. This literature review shows a richness of work in recent years that describe the development of various systems using deep learning to solve various problems in precision agriculture. The problems addressed are mainly yielded losses caused by factors affecting crop growth. In bibliometric methodology, we explored only documents from the Web of Science database. We analyzed the thematic structures and new directions of agricultural research based on deep neural networks were proposed.

     Bibliometric analyses show that the most active countries are China, the United States, India, and Brazil. The hot topics covered include plant disease, feature extraction, transfer learning, disease detection, weed detection, insect pests, convolutional neural networks, image processing, remote sensing, and IoT. Convolutional neural networks (CNNs) are one of the most used architectures thanks to their successful application in deep learning. However, the use of deep learning algorithms in precision agriculture raises many challenges whose resolution is useful to enhance this field and facilitate its appropriation by the stakeholders: