Efficacy of machine learning, earth observation and geomorphometry for mapping salt-affected soils in irrigation fields

Date
2018-12
Journal Title
Journal ISSN
Volume Title
Publisher
Stellenbosch : Stellenbosch University
Abstract
ENGLISH ABSTRACT: There is a need to monitor salt accumulation throughout agricultural irrigation schemes as it can have a major negative impact on crop yields and subsequently result in a lower food production. Salt accumulation can result from natural processes, human interference or prolonged waterlogging. Most irrigation schemes are large and therefore difficult to monitor via conventional methods (e.g. regular field visits). More cost-effective, less time-consuming approaches in identifying salt-affected and salt-prone areas in large irrigation schemes are therefore needed. Remote sensing has been proposed as an alternative approach due to its ability to cover a large region on a timely basis. The approach is also more cost-effective because less field visits are required. A literature review on salt accumulation and remote sensing identified several direct and indirect methods for identifying salt-affected or salt-prone areas. Direct methods focus on the delineation of salt crusts visible on the bare soil in multispectral satellite imagery, whereas indirect methods, which include vegetation stress monitoring and geomorphometry (terrain analysis), attempt to take subsurface conditions into account. A disadvantage of the direct approach is that it does not take subsurface conditions into account, while vegetation stress monitoring (an indirect method) can produce inaccurate results because the vegetation stress can be a result of other factors (e.g. poor farming practices). Geomorphometry offers an alternative (modelling) approach that can either replace or augment direct and other indirect methods. Two experiments were carried out in this study, both of which focussed on machine learning (ML) algorithms (namely k-nearest neighbour (kNN), support vector machine (SVM), decision tree (DT) and random forest (RF)) and statistical analyses (regression or geostatistics) to identify salt-affected soils. The first experiment made use of very high resolution WorldView-2 (WV2) imagery. A number of texture measures and salinity indices were derived from the WV2 bands and considered as predictor variables. In addition to the ML and statistical analyses, a classification and regression tree (CART) model and Jeffries-Matusita (JM) distance thresholds were also produced from the predictors. The CART model was the most accurate in differentiating salt-affected and unaffected soils, but the accuracy of kNN and RF classifications were only marginally lower. The normalized difference salinity index showed the most promise among the predictors as it featured in the best JM, regression and CART models. The second experiment applied geomorphometry approaches to two South African irrigation schemes. Elevation sources include the Shuttle Radar Topographic Mission (SRTM) digital elevation model (DEM) and a digital surface model (DSM) produced from stereoscopic aerial photography. A number of morphological (e.g. slope gradient) and hydrological (e.g. flow direction) terrain parameters were derived from the SRTM DEM and the DSM and used as predictors. In addition to the algorithms used for the first experiment, the geostatistical method Kriging with external drift (KED) was also evaluated in this experiment. The source of elevation had an insignificant impact on the accuracies, although the DSM did show promise when combined with ML. KED outperformed regression modelling and ML in most cases, but ML produced similar results for one of the study areas. The experiments showed that direct and geomorphometry approaches hold much potential for mapping salt-affected soil. ML also proved to be a viable option for identifying salt-affected or salt-prone soil. It is recommended that a combination of direct and indirect (e.g. vegetation stress monitoring) approaches are considered in future research. Making use of alternative data sources such as hyperspectral imagery or higher spatial resolution DEMs may also prove useful. Clearly, more research is needed before such approaches can be operationalized for detecting, monitoring and mapping salt accumulation in irrigated areas.
AFRIKAANSE OPSOMMING: Daar is 'n behoefte om soutophoping deur middel van landboubesproeiingskemas te monitor aangesien dit 'n beduidende negatiewe uitwerking op oesopbrengste kan hê en gevolglik tot laer voedselproduksie kan lei. Soutophoping kan voortspruit uit natuurlike prosesse, menslike inmenging of langdurige deurdrenking. Die meeste besproeiingskemas is groot en daarom moeilik om te monitor via konvensionele metodes (bv. gereelde veldbesoeke). Meer koste-effektiewe, minder tydrowende benaderings is dus nodig om soutgeaffekteerde areas en areas wat geneig is tot soutophoping in groot besproeiingskemas te identifiseer. Afstandswaarneming is voorgestel as 'n alternatiewe benadering weens sy vermoë om 'n groot streek op 'n tydige basis te dek. Die benadering is ook meer koste-effektief omdat minder veldbesoeke vereis word. 'n Literatuuroorsig oor soutophoping en afstandswaarneming het verskeie direkte en indirekte metodes geïdentifiseer om soutgeaffekteerde areas of areas geneig tot soutophoping te identifiseer. Direkte metodes fokus op die afbakening van soutkorste wat in multispektrale satellietbeelde op die kaal grond sigbaar is. Indirekte metodes, insluitende plantstresmonitering en geomorfometrie (terreinanalise), aan die ander kant, poog om die ondergrondse toestande in ag te neem. 'n Nadeel van die direkte benadering is dat dit nie ondergrondse toestande in ag neem nie, terwyl plantstresmonitering ('n indirekte metode) onakkurate resultate kan veroorsaak, aangesien die plantstres die gevolg kan wees van ander faktore (bv. swak boerderypraktyke). Geomorfometrie bied 'n alternatiewe (modellering) benadering wat direkte of ander indirekte metodes kan vervang of uitbrei. In hierdie studie is twee eksperimente uitgevoer. Albei het gefokus op masjienleer (ML) algoritmes, naamlik k-nearest neighour (kNN), ondersteunende vektormasjien, besluitboom en ewekansige woud (EW), en statistiese ontledings (regressie of geostatistiek) om soutgeaffekteerde gronde te identifiseer. Die eerste eksperiment het gebruik gemaak van baie hoë resolusie WorldView-2 (WV2) beelde. 'n Aantal tekstuurmaatreëls en soutindekse is afgelei van die WV2-bande en is beskou as voorspeller-veranderlikes. Benewens die ML en statistiese ontledings, is 'n klassifikasie- en regressieboom (KARB) model en Jeffries-Matusita (JM) afstandsdrempels ook van die voorspellers vervaardig. Die KARB-model het die mees akkuraatste differensiasie tussen sout-geaffekteerde en ongeaffekteerde grond gemaak, maar die akkuraatheid van kNN- en EW-klassifikasies was slegs marginaal laer. Van al die voorspellers het die genormaliseerde-verskil-saliniteit-indeks die meeste belofte getoon aangesien dit in die beste JM-, regressie- en KARB-modelle presteer het. Stellenbosch University https://scholar.sun.ac.za vi Die tweede eksperiment het geomorfometriese benaderings toegepas op twee Suid-Afrikaanse besproeiingskemas. Elevasiebronne sluit in die Shuttle Radar Topographic Mission (SRTM) digitale elevasie-model (DEM) en 'n digitale oppervlakmodel (DOM) wat uit stereoskopiese lugfotografie vervaardig word. 'n Aantal morfologiese (bv. hellingsgradiënt) en hidrologiese (bv. vloeirigting) terreinparameters is afgelei van die SRTM DEM en die DOM en is gebruik as voorspellers. Benewens die algoritmes wat vir die eerste eksperiment gebruik is, is die geostatistiese metode Kriging met eksterne dryf (KED) ook in hierdie eksperiment geëvalueer. Die bron van elevasie het 'n onbeduidende impak op die akkuraatheid gehad, hoewel die DOM belofte getoon het wanneer dit met ML gekombineer is. KED het in meeste gevalle beter presteer as regressie modellering en ML, maar ML het soortgelyke resultate vir een van die studiegebiede opgelewer. Die eksperimente het getoon dat direkte en geomorfometriese benaderings baie potensiaal het vir die kartering van soutgeaffekteerde grond. ML het ook bewys dat dit 'n lewensvatbare opsie is om soutgeaffekteerde grond of grond wat geneig is tot soutophoping, te identifiseer. Daar word aanbeveel dat 'n kombinasie van direkte en indirekte (bv. plantegroei-stresmonitering) benaderings in toekomstige navorsing oorweeg word. Die gebruik van alternatiewe databronne soos hiperspektrale beelde of hoër ruimtelike resolusie-DOM's kan ook nuttig wees. Dit is duidelik dat meer navorsing nodig is voordat sulke benaderings geoperasionaliseer kan word vir die opsporing, monitering en kartering van soutophoping in besproeide gebiede.
Description
Thesis (MSc)--Stellenbosch University, 2018.
Keywords
Earth sciences -- Remote sensing -- Field irrigation, Soil management -- Geology -- Statistical methods, Machine learning -- Geology -- Statistical methods, Geomorphometry -- Soils, Salts in -- Field irrigation, UCTD
Citation