Colloquium

Mapping cashew plantations in Côte d’Ivoire: A comparative study of open source spatial data and machine learning techniques to map cashew (plantations) in all of Côte d’Ivoire

Organised by Laboratory of Geo-information Science and Remote Sensing
Date

Tue 28 March 2023 09:00 to 09:30

Venue Gaia, building number 101
Droevendaalsesteeg 3
101
6708 PB Wageningen
+31 (0) 317 - 48 17 00
Room 1

By Mark Boeve

Abstract
Côte d’Ivoire has recently become the largest producer of cashew nuts in the world. For 1.5 million people in Côte d’Ivoire, cashew nuts are the main source of income. However, cashew nut production and the expected halt in its proceeds (are expected to) have major impact on the citizens and the environment in Côte d’Ivoire. Information about the total surface of cashew plantations is lacking, while its value for policy makers, conservationists and others has increased substantially. The objective of this study was to map all cashew plantations in Côte d’Ivoire with the use of Remote Sensing and machine learning. Four classifiers (Classification And Regression Trees (CART), Gaussian Naive Bayes (GNB), Random Forest (RF), and Support Vector Machine (SVM) classifier) were compared in combination with four open (except for PlanetScope) spatial data sources (Landsat-8, Sentinel-2, PlanetScope, and Sentinel-1) and one elevation data set (SRTM 90m DEM). It was found that Sentinel-2 data and the RF-classifier are most suitable for classifying cashew nut plantations in Côte d’Ivoire. The RF-classifier is the most accurate and has a relatively low computation time. Classifications of all of Côte d’Ivoire executed with Sentinel-2 data were slightly more accurate than classifications executed with Landsat-8 data. But, classifications executed with Landsat-8 data had a substantially lower computation time. Classifications executed with PlanetScope or Sentinel-1 data were significantly less accurate. Besides, it was found that adding elevation or Sentinel-1 data to the classification data improved the accuracy of the classifications scantly, whereby it was stated negligible. The result of this study is a classification map of cashew plantations in the north of Côte d’Ivoire. The south could not be classified accurately due to almost permanent cloud cover. For further research, it is therefore recommended to improve the pre-processing of Sentinel-1 data to increase the amount of available data for classifying Côte d’Ivoire.

Keywords: cashew plantations; remote sensing; classification; machine learning; Côte d’Ivoire; Classification And Regression Trees; Gaussian Naive Bayes; Random Forest; Support Vector Machine; Landsat-8; Sentinel-2; PlanetScope; Sentinel-1; SRTM DEM