Prediction of desert locust breeding areas using machine learning methods and SMOS (MIR_SMNRT2) Near Real Time product

Abstract

Despite satellite imagery is being used to identify suitable areas for desert locust, there is a lack of automatized and operational procedures in Near Real Time (NRT). The aim of this study was to assess the capacity of Soil Moisture Near Real Time Neural Network Level 2 product (MIR_SMNRT2) from the Soil Moisture and Ocean Salinity satellite (SMOS) to predict nymphs of desert locust. We used soil moisture time series (between 2016 and 2019) to build 6 machine learning models (logistic regression model “glm”, eXtreme Gradient Boosting “xgbTree”, Weighted k-Nearest Neighbors “kknn”, Feed-Forward Neural Networks and Multinomial Log-Linear Models “nnet”, support vector machine radial “svmRadial”, and random forest “rf”) over the entire recession area. Model results proved that spatial and/or temporal constraints in data sampling conditioned the predictive capacity of the selected machine learning algorithms. Furthermore, we used a forward selection procedure to evaluate the impact that time series data exert on modelling. Our results suggest that soil moisture data retrieved between 95 and 12 days (before the sighting) provided sufficient information to achieve acceptable predictive performances. This methodology can improve current preventive and control operations, it is site-specific, and could be used to other pests.