A machine learning system for IoT control of irrigation and fertilization to optimize rice yield in Rwanda

Abstract

Water and fertilization are widely recognized as essentials for optimal rice plant growth. Efficient use of water and fertilization for agriculture are critical to ensure high yields and maximize economic benefits. The central role of water access for agriculture is a clear challenge everywhere in Rwanda, especially in areas with significant seasonal variation in rainfall such as Muvumba North east Rwanda. it fails to increase the resilience of agricultural systems in the face of complex demands for water use in rice due to each stage needs its amount of water in dependent to previous one. In Muvumba, where the farmers have a low level of economic development are facing the problem of infrastructure, lack of irrigation control for individual farmers, lack of access to equipment, and low reliability of power and Internet access. Applying IoT technology will solve the problem that is why in our thesis explores al gorithms using Markov chain process that automatically provide irrigation control ac cording to the stage of rice, when the system are operating correctly. In cases of system component failure, the system switches to an alternative prediction mode called SARSA. The SARSA algorithm outputs realistic irrigation options depending on previous data from Markov chain process algorithm until the failure is corrected. Farmers can receive information about the faults and suggested actions via SMS. Both algorithms are exam ined using simulations to assess how the system might respond to growth stage, effective rainfall, and evapotranspiration for both correct operation and failure scenarios. Regarding fertilization, Muvumba plantation suffers from poor fertilization management due to only one laboratory for testing soil nutrients which causing delays in soil testing and information dissemination. Here, two algorithms based on fuzzy logic were designed xvi xxxi with input from well-known best practices for local conditions. The first is a nutrient balance method for automatic decision making and the second is dissimilar subtraction for the case of system fault. The fuzzy algorithms have a linguistic rule base of 183 IF THEN statements linking measurable field conditions to crop yield. These rules were designed using input from interviews with Government of Rwanda (GoR) agricultural experts and published knowledge of site conditions. The rules incorporate the known nutrient requirements of the different growth stages of rice. To validate the algorithms, historical weather and field data are used to drive simulations of yield for different plots during the season A(September-march) of 2020 at sites in Northeast Rwanda. Predicted yields are compared to measured yields for scenarios with different irrigation levels and fertilization amounts and with and without full Internet connectivity.In case of fault tol erance in the commonly occurring case of network communications failure, an dissimilar subtraction algorithm where the farmer is informed on the system status and recom mended actions via SMS through a GSM. The novelty of our work lies on designing low-cost IoT algorithms system would au tomatically provide irrigation and fertilization control according to seasonal and daily irrigation or fertilization needs when the system sensors and communications are operat ing correctly. In cases of system component failure, the system switches to an alternative prediction mode and messages farmers with information about the faults and realistic irrigation or fertilization options until the failure is corrected controls water and fertilizer on each stage of rice more efficiently with fault tolerance to optimize yields.