Comparative Performance Analysis of Machine Learning and Deep Learning Models for IoT-Based Crop Yield Prediction

Abstract

The environmental uncertainty and the nonlinear behaviour of IoT-sensor data nature influence the crop yield prediction. In Nawab Shah, this region requires a model which behaves capturing complex patterns. This study provides a deep learning and machine learning comparative analysis, whereas data is obtained from the IoT-based sensor data; the dataset parameters include temperature, humidity, smoke, and light intensity. The dataset has more than 35000 time-stamped samples, collected with the five seconds delay in each data reading.  These observations are processed by applying data normalisation, outlier detection and cleaning, and min-max normalisation. The statistical data validation is obtained by applying RMSE, MAE, MSE, Pearson’s correlation and confusion matrix. The Bayesian-optimised random forest consistently achieved outstanding performance with the highest accuracy, recall, and precision with 0.33 F1-Score. The smoke and humidity are the significance factor from the obtained results analysis for the yield prediction. The classification ability is confirmed by the confusion matrix with the ability as average, good and poor classes of the yield. Furthermore, the finding shows that the optimised random forest performed better than all in the environmental data for the prediction of yield. This is also based on the same features as smoke and humidity; this methodology and approach provide a reliable and low-complex framework with a real-time precision agriculture system for decision-making. LSTM models, along with a variant of Random Forest, give results of 0.27 intermediate range value, which recommends that the very important patterns are captured, but are not effective for the top models.