Assessing Soil and Land Suitability of an Olive–Maize Agroforestry System Using Machine Learning Algorithms

Abstract

Exponential population increases are threatening food security, particularly in mountainous areas. One potential solution is dual-use intercropped agroforestry systems such as olive (Olea europaea)–maize (Zea mays), which may mitigate risk by providing multiple market sources (oil and grain) for smallholder producers. Several studies have conducted integrated agroforestry land suitability analyses; however, few studies have used machine learning (ML) algorithms to evaluate multiple variables (i.e., soil physicochemical properties and climatic and topographic data) for the selection of suitable rainfed sites in mountainous terrain systems. The goal of this study is therefore to identify suitable land classes for an integrated olive–maize agroforestry system based on the Food and Agriculture Organization (FAO) land suitability assessment framework for 1757 km2 in Khyber Pakhtunkhwa province, Pakistan. Information on soil physical and chemical properties was obtained from 701 soil samples, along with climatic and topographic data. After determination of land suitability classes for an integrated olive–maize-crop agroforestry system, the region was then mapped through ML algorithms using random forest (RF) and support vector machine (SVM), as well as using traditional techniques of weighted overlay (WOL). Land suitability classes predicted by ML techniques varied greatly. For example, the S1 area (highly suitable) classified through RF was 9%↑ than that of SVM, and 8%↓ than that through WOL. The area of S2 (moderately suitable) classified through RF was 18%↑ than that of SWM and was 17%↓ than the area classified through WOL; similarly, the S3 (marginally suitable) class area via RF was 27%↓ than that of SVM, and 45%↓ than the area classified through WOL. Conversely, the area of N2 (permanently not suitable class) classified through RF and SVM was 6%↑ than the area classified through WOL. Model performance was assessed through overall accuracy and Kappa Index and indicated that RF performed better than SVM and WOL. Crop suitability limitations of the study area included high elevation, slope, pH, and large gravel content. Results can be used for sustainable intensification in mountainous rainfed regions by expanding intercrop agroforestry systems in developing nations to close yield gaps.

Keywords

• random forest
• support vector machine
• intercropping agroforestry
• land suitability