Towards precision agriculture in Morocco: A machine learning approach for recommending crops and forecasting weather

Abstract

Statistical models predict that the world's population will reach 8.5 billion by the end of 2030. This represents a real threat to our food security and puts the current food production system under pressure. Efficient use of Earth's natural resources is the only solution to facing future challenges such as global hunger. The implementation of precision agriculture using new technologies such as artificial intelligence, big data, IoT and remote sensing is the first step towards this goal. In this paper, we investigated several machine learning models to create two services: one for recommending the best crop to grow based on soil and the region's weather characteristics, and another for the forecasting of the hourly average air temperature. Performance evaluation results for the first service show that Random Forest has the best metrics as a classifier (accuracy = 100%, precision = 100%, recall = 100%) compared to K-Nearest Neighbors (KNN), Decision Tree, Naive Bayes, Logistic Regression, Convolutional Neural Network, and Feed Forward Neural Network. This is a confirmation that classic machine learning algorithms perform better on small-size datasets. In our case, we used a dataset of 2200 instances available online. On the other hand, Facebook Prophet was more accurate (R2 = 0.81, RMSE = 3.74) than our proposed LSTM architecture in time series forecasting at hourly scale using historical weather data provided by the weather station of our study area. These two optimal models are then integrated as the first building blocks in our decision support platform, intended for both farmers and policymakers with the aim of making agriculture in Morocco more efficient and more sustainable.