A Genetic Programming Approach to Estimate Vegetation Cover in the Context of Soil Erosion Assessment

Abstract

This work describes a genetic programming (GP) approach that creates vegetation indices (VI's) to automatically detect the sum of healthy, dry, and dead vegetation. Nowadays, it is acknowledged that VI's are the most popular method for extracting vegetation information from satellite imagery. In particular, erosion models like the "Revised Universal Soil Loss Equation" (RUSLE) can use VI's as input to measure the effects of the RUSLE soil cover factor (C). However, the results are generally incomplete, because most indices recognize only healthy vegetation. The aim of this study is to devise a novel approach for designing new VI's that are bettercorrelated with C, using field and satellite information. Our approach consists on stating the problem in terms of optimization through GP learning, building novel indices by iteratively recombining a set of numerical operators and spectral channels until the best composite operator is found. Experimental results illustrate the efficiency and reliability of our approach in contrast with traditional indices like those of the NDVI and SAVI family. This study provides evidence that similar problems related to soil erosion assessment could be analyzed with our proposed methodology.

Evolutionary-computer-assisted design of image operators that detect interest points using genetic programming

“…For these reasons, researchers in other fields might become tentative, or even skeptical, of solutions that are generated by GP. On the other hand, the works from the first group, those that attempt to detect features defined by human experts, by definition will not be hampered by the problem of semantic interpretation and here we can find examples also for object detection [43,44], and the analysis of multi-spectral images [47]. Furthermore, we believe that when appropriate fitness criteria are given, and when a comprehensive analysis of the obtained results is carried out, then it is possible to derive a better understanding of the logic behind the solutions that a GP produces, and also to obtain deeper insights regarding the nature of the problem itself.…”

“…It is possible to identify three types of GP-based approaches: (1) those that employ GP to detect low-level features which have been predefined by human experts, such as corners or edges [21,44,[60][61][62]67] and recently one regarding vegetation indices used on remote sensing [46,47];…”

Mapping Ecological Processes and Ecosystem Services for Prioritizing Restoration Efforts in a Semi-arid Mediterranean River Basin

“…For the extrapolation of vegetation cover density, we used the cover factor of the RUSLE model (Renard et al 1997) called C factor, which is the cover-management term, with the subfactors consisting of prior land use, crop canopy and surface cover . Trabucchi et al (2012) applied to the study area the model using Genetic Programming Vegetation Index (GPVI) (Puente et al 2011) to extract the cover factor, which evaluates vegetation land cover ranging from 0.45-1 and is part of the RUSLE model formula.…”

Synthesis of Vegetation Indices Using Genetic Programming for Soil Erosion Estimation

“…In the literature reviewed for this article, it was found that the only machine learning studies that have tried to use the synthesis of vegetation indices for the C factor are the works by Puente et al [29] and Trabucchi et al [30]. In [29], a first approximation to the synthesis of vegetation indices using Genetic Programming (GP) for the Todos Santos basin was reported with positive results; while [30] showed how synthesized vegetation indices are able to identify areas prone to erosion in the Rio Martin basin. Vegetation indices and specifically NDVI have been widely used in studies dealing with landslide [31], susceptibility to soil erosion [32,33], and gully erosion [34] There are two main reasons for applying GP for synthesizing the C factor.…”

“…These cover the visible and infrared electromagnetic spectrums and have a resolution of 30 × 30 m per pixel. Band 6, thermal infrared, is not considered relevant for this type of study [15,26,[28][29][30]32]. Therefore, we decided not to use it.…”