Design and Implementation of a Deep Learning Solution for Soil Moisture Prediction Using Local Farm Meteorological Data
Project
Soil moisture is one of the fundamental factors in agricultural production and hydrological cycles. Its precise prediction is crucial for the rational use and management of water resources. However, soil moisture involves complex structural characteristics and wheather factors, and it is difficult to establish an ideal mathematical model for predicting soil moisture. Existing predictive models present problems such as predictive accuracy, generalization and processing capacity of several characteristics, and predictive performance needs to be improved. Being a company of intelligent solutions in agriculture, the startup Viable Ways Corp wanted to respond to this problem and implement a solution for the prediction of soil moisture. This report summarizes the work done during my graduation project at the School of Information Science (ESI) for the State Engineering degree in Data and Knowledge Engineering, That I carried out within Viable Ways Corp. The objective of this work is to predict soil moisture based on weather data, looking for the most suitable model for our case.
Survey of existing
The company is undertaking a large-scale IoT project, installing smart sensors on farms to measure real-time values of physical parameters and transmitting them to the Cloud for storage. Subsequently, Viable Ways Corp is engaged in leveraging these data through artificial intelligence tools to develop decision support systems.
The technological solutions proposed by the system will gather data on the farm's GPS location, forecasted precipitation and weather conditions, soil type, and nutrition. The system will process information through smart cloud analytics to assist farmers in making better agricultural decisions. Farmers will receive mobile notifications providing insights on optimal crop choices, the optimal planting week, pest growth alerts, weather advisories, market information, and agriculture tips developed in collaboration with other partners.
The company doesn't have a soil moisture forecasting and prediction system. it offers KPI analysis technologies so the farmer can access his weather conditions in real time, however, he doesn't have a clear forecast of what's going to happen.
Situation
Solution
Results
To achieve our goal, I adopted several approaches: a regression approach using RFR, a prediction approach using LSTM and ARIMA.
The table alongside shows the models used, whose performance was tested on the basis of MAE, RMSE and r2. The best performances were observed with the ARIMA and RFR models. However, we have to take into consideration that the ARIMA model, can only predict the next 48 hours, which is interesting in model terms, but may not be considerable enough in terms of soil moisture prediction.
On the other hand, the LSTM model gave good results for a longer duration. The RFR model proved its performance in terms of accuracy and forecast duration.
Finally, we can conclude that the model most adaptable to our case is the Random Forest Regressor with a forecast duration of six days and an accuracy of 94%. ( For in-depth tech insights, hit me up on LinkedIn. Let's dive deeper into this research together!)
