The paper presents a study on the development and integration of a recurrent neural network (RNN) to improve the accuracy and fault tolerance of a moisture transfer model in a smart garden system. The problem of soil moisture control is becoming especially relevant in modern agricultural and environmental monitoring, where high accuracy is required to manage water resources, forecast crop yields and prevent drought periods. Traditional methods, such as remote sensing and moisture transfer models, have significant limitations: low accuracy, computational complexity, dependence on accurate sensor data and difficulty in applying in real field conditions. To solve these problems, the study proposes the use of RNN, which is able to effectively process time series data and predict soil moisture even in the presence of incomplete, inaccurate or distorted input data. The global soil moisture dataset GSSM and weather data from the Meteostat platform were used as initial data, which made it possible to take into account the climatic features of regions with different soil types. The model includes a long short-term memory (LSTM) layer and a fully connected layer for the final forecast. Particular attention is paid to data preprocessing, including calculating average daily, average monthly and average annual values, as well as data correction taking into account the characteristics of different soil types. The study showed that the developed RNN model is highly resistant to sensor failures, has minimal dependence on the volume of input data and is able to adapt to different climatic and soil conditions. The proposed solution improves the accuracy of soil moisture monitoring in the Smart Garden system, optimizes the use of water resources and increases the stability of the system in the face of changing external factors. Thus, the integration of RNN opens up new opportunities for the development of agriculture and ecology, ensuring more efficient water resource management and increasing the productivity of agrosystems
