Comparison of Multiple-Layer Perceptrons and Least Squares Support Vector Machines for Remote-Sensed Characterization of In-Field LAI Patterns – A Case Study with Potato

Abstract

Delineation of soil management zones in agricultural fields using reliable indicators is a major issue in precision agriculture. The leaf area index (LAI) is an important variable for the characterization of in-field variability. However, ground LAI measurement over large fields is time consuming. Our objective was to compare machine learning methods to describe in-field potato LAI patterns from airborne multispectral images. To this aim, intensive ground LAI measurements (97 quadrats) were collected in a potato field at the time of maximum LAI. Two methods were trained as function approximation, validated, and compared to linear regressions. The two methods were (i) multiple-layer perceptron (MLP) and (ii) least squares support vector machine (LS-SVM). After model training, spatial interpolation was performed and results were compared to a map interpolated with measured values. Both methods performed well using near-infrared and red channels as inputs. However, the gain in performance in validation over the best linear model was higher for the LS-SVM (29%) compared to the MLP (15%), and the kappa coefficient of agreement was higher during classification. The LS-SVM with 2 inputs (near-infrared and red) was therefore retained as the final model.