Remote sensing and machine learning techniques applied to gully mapping: a systematic review

Abstract

The development of gully inventory and susceptibility maps, obtaining measures of their extent and the assessment of related environmental factors are crucial for quantifying and characterizing land degradation, along with supporting recovering actions and dealing with erosion impacts. Different approaches may be applied to produce gully maps, including field-based investigations, visual interpretation, pixel- and object-based methods. A systematic review was conducted using PRISMA guidelines and Google Sheets to assess gaps and recent advances in selected publications regarding the research question ‘What are the leading artificial intelligence methods for automatic gully mapping and how is their performance?’ Then, among 1951 records found, 186 publications were selected and screened. The review addressed the differences and performances of imagery and topographic input, gully inventory mapping (pixel- or object-based methods), gully susceptibility mapping and associated algorithms, and accuracy assessment. Hence, we discussed the advancements in gully erosion mapping techniques, including inventory and susceptibility, and the comparative assessment of imagery and topographic inputs. We highlight the use of Google Earth, UAV and Sentinel-2 imagery, along with ALOS PALSAR DEM, SRTM, UAV DEM and ASTER DEM topographic input. The Random Forest algorithm showed consistent results in performance both in pixel- and object-based methods for gully inventory and susceptibility mapping, although ensemble models exhibited better accuracy, in some cases. Finally, the object-based techniques demonstrated higher suitability for remotely assessing and mapping gullies due to the application of segmentation of objects before the machine learning modelling.