Smart farming (SF) is a sustainable farm management concept used for the intensification of food production to meet the growing demand. With the progress of the Internet of Things, advanced systems have been widely proposed for monitoring and processing data to generate insights that help producers to optimize farm management processes. Centralizing data to a remote Cloud is the conventional data processing method, though the extended latency in getting insights back to the application and intermittent Internet connectivity limit its adoption particularly in time-sensitive applications. Alternatively, distributed data analytics methods have been introduced to enable processing data in proximity to the sources and then combine insights accordingly for making timely and accurate decisions cooperatively. However, most of the SF systems currently in use operate in isolation due mainly to the lack of analytic techniques that can effectively incorporate them for processing data. Consequently, their full potential as well as the data collected by them is significantly under-utilized. This PhD research focuses on the development of distributed data processing and learning methods to enable cooperative data analytics. Initially, this research explores how large scale complex data can be simplified for conducting effective analysis and then proposes a Compressed Learning (CL) approach and a novel metric, known as animal importance (AIm), to extract meaningful information to perform learning effectively. To illustrate the potential of the CL approach in processing large-scale data in the SF domain, this study presents an application of CL in analyzing large-scale Mid-Infrared (MIR) milk quality data. Also, as an application of the AIm metric in the smart dairy farming domain, the research discuses how effectively AIm could be used for alerting the prevalence of sick and estrus cows in a herd based on the variability in behavioral dynamics. Second, this PhD research develops a hybrid model to mitigate drawbacks that limit using conventional machine learning models and proposes the Federated Learning (FL) method to train distributed data sources cooperatively. The FL-based system is analyzed to determine its applicability for assessing milk quality by incorporating MIR milk quality data collected at distributed farms. This is then followed by considering the fact the limitations of the FL-based approach when it comes to making the data analytics more trustable and transparent to every participant in the distributed network, by integrating a Block Chain-enabled fully decentralized distributed learning framework. In particular, this framework integrates the Internet of Nano Things (IoNT) that has previously not taken into account any Block Chain-enabled system. The proposed framework is then used for monitoring the level of chemicals (e.g., fertilizers) on farmlands. Finally, this PhD research discusses optimum utilization of available resources in cooperative distributed data analytics by offloading computations to neighboring devices. Computation offloading enhances the timeliness and learning accuracy in cooperative data analytics as well as enabling the efficient use of limited energy resources found in sensor devices, and this includes solar energy harvesting devices.
|Publication status||Unpublished - 2019|
- Distributed learning, Data processing, Smart Farming