Cotton Jassid Infestation Monitoring Using Sentinel-2 MSI, Landsat 9 TIRS, and PlanetScope Imagery with Machine Learning and Data Fusion

Abstract

The excessive use of synthetic pesticides in the fields is the cause of numerous problems to human health and biodiversity. Knowing the location and severity of the emergence of pests enables targeted treatment, helping to reduce the use of pesticides. Given that traditional field inspection requires considerable effort with limited effectiveness, remote sensing combined with machine learning is an essential complement for large-scale, efficient monitoring. This study therefore aims to develop a machine learning-based solution for detecting cotton jassids from free satellite imagery data in Alibori, northern Benin. Multispectral and thermal infrared images respectively from the Sentinel-2 and Landsat 9 satellites were fused in the most correlated bands using signal processing algorithms Pseudo Wigner Distribution (PWD), Nonsubsampled Contourlet Transform (NSCT), and Discrete wavelet transform (DWT). Two configurations, in terms of data augmentation, were used. Areas infested by jassids were identified in merged and unmerged satellite images (Sentinel-2, Landsat 9, and PlanetScope, a commercial satellite) using Support Vector Machine (SVM), Random Forest (RF), and CatBoost supervised machine learning algorithms. The green and Thermal Infrared Sensor 1 (TIRS1) bands are the most correlated in configuration 1 and the blue and TIRSI bands in configuration 2 for respectively Sentinel-2 and Landsat 9. The DWT and NSCT signal processing algorithms produced the least Root Mean Square Deviation (RMSE) for image fusion (0.46 for DWT and 0.33 for NSCT). The best result in terms of detecting performance was obtained with the composite treatment of Landsat 9 images and the RF algorithm, with an overall accuracy of 88.23%, an Fl-score of 82%, and a Kappa index of 79.92% in configuration 1 and in configuration 2 for composite treatment of Sentinel-2 an overall accuracy of 92.36%, an Fl-score of 90.75<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">%</sup> and a Kappa index of 86.36% with CatBoost algorithm.