Prediction Of Soil Orders With High Spatial Resolution . - UFRGS

Prediction of soil orders with high‑spatial resolutionarea contains 155 polygons and 37 individual soil typesbelonging to four soil orders: 10 Argissolos (Ultisols), 16Cambissolos (Inceptisols), 4 Chernossolos (Mollisols),and 7 Neossolos (Entisols). Calculation of predictorvariables, spatial analysis, prediction of soil classes,and accuracy assessment were done using the softwareIdrisi Taiga GIS (Clark Labs, Worcester, MA, USA).The first step was the selection of prediction variables.Based on local expert knowledge of soil formationfactors and on data availability, variables correlatedwith variations on moisture regime, erosion anddeposition of sediments, organic matter concentration,and depth of the A horizon were considered. Some ofthe soil formation factors are uniform throughout thestudy area, including major geology units and climate –particularly high annual rainfall –, whereas others, suchas land cover, were not mapped on the spatial resolutionused in the present work. However, microclimaticvariables and land use are strongly conditioned byrelief. In fact, the strong influence of relief on soilformation in the evaluated area is well known (Floreset al., 1999), indicating that terrain variables should begood predictors of soil classes (Florinsky et al., 2002).According to Giasson et al. (2011), variables that candescribe these variations in the region are: elevation,slope, aspect, profile curvature, flow accumulation,flow direction, and planar distance from streams. Thefirst six predictors were calculated directly from theDEM, and the last variable was calculated from thestream network.Generation of sampling points was done with randomspatial distribution, using five sampling densities,comprised in the recommended range for detailedsoil surveys in Brazil (Manual técnico de pedologia,2007): 0.5, 1, 1.5, 2, and 4 points per hectare. Sincesoil classification depends on a number of physicaland chemical characteristics, before sampling, theconventional soil map was simplified so that theclasses stayed coherent with the selected predictor1397variables and could be properly sampled. Soil typeswere grouped to the first taxonomic level (order),resulting in a soil map with four classes: Argissolos,Cambissolos, Chernossolos, and Neossolos. Then,values of predictor variables and soil order, at eachsampling point, were collected for all samplingdensities. Grouped soil classes and number of samplepoints per class are shown in Table 1.Data from sampling points were used to train mLAand to predict the occurrence of soil orders in the wholestudy area. Four classification algorithms, based onthe concept of mLA, were used: three artificial neuralnetworks (multi‑layer perceptron, MLP; adaptiveresonance theory, fuzzy ARTMap; and self‑organizingmap, SOM) and a decision tree (Gini). Artificialneural networks simulate the operation of the structureof neurons and connections of the human brain,whereas decision trees simulate the human process ofabstraction through hierarchical categorization (Lippittet al., 2008). In the training process, 10,000 iterationswere used, aiming to optimize the algorithm’s structureand to reach stability on prediction error.For accuracy assessment, each predicted soil mapwas compared with the conventional soil map, usingall pixels of the study area to calculate error matrices(Congalton, 1991), and to compute five accuracyindicators: omission errors, expressed as the proportionof a specific class that was estimated as other classes;commission errors, expressed as the proportion ofdifferent classes included in a specific estimatedclass; overall accuracy, expressed as the proportionof correctly‑classified pixels; quantity disagreement,which measures the amount of difference betweenthe reference map and the estimated map attributedto the less than perfect match in the proportions of thecategories; and allocation disagreement, which measuresthe amount of difference between the reference mapand the estimated map due to the less than optimalmatch in the spatial allocation of the categories, givenTable 1. Classes of the grouped soil map (order), according to the Brazilian soil classification system (SiBCS) and to SoilTaxonomy, and area, proportion, and number of sample points per class at each sampling solosTotalSoil rea (ha)101.6281.3228.761.9100.0Proportion oints per 036888233Pesq. agropec. bras., Brasília, v.47, n.9, p.1395-1403, set. 2012

1398E.C. Sarmento et al.the proportions of the categories in the reference andestimated map.Quantity disagreement and allocation disagreementwere preferred instead of kappa, which, accordingto Pontius & Millones (2011), provides redundantinformation and does not give guidance on how toimprove classification. While kappa measures howmuch the agreement is better than random, quantitydisagreement and allocation disagreement measurehow much the agreement is less than perfect, providingadditional information that helps to explain error.The MLP neural network, with 0.5 point perhectare, simultaneously showed the lowest omissionerror for Cambissolos and the highest omission errorfor Neossolos and Argissolos (Table 2). However, itpresented minimum commission error for Neossolosand Argissolos (Table 3), since the algorithm could notestimate these classes. In this case, the omission error ismaximized and the commission error is minimized. Atthe same time, predicted classes that incorrectly receivepixels from unpredicted classes have their omissionerror reduced and their commission error increased(Congalton, 1991; Pontius & Millones, 2011). Thevery low omission error observed for Cambissolosindicates that most of the omitted pixels of Neossolosand Argissolos were incorrectly allocated to that class.Considering only the cases in which all classes couldbe estimated, the lowest omission errors for Argissolosand Neossolos were found using the Gini decisiontree with sampling density of four points per hectare.Lowest errors for Cambissolos and Chernossolos wereobserved using the MLP neural network, with 1.5 and4 points per hectare, respectively. The Gini decisiontree and MLP neural network also showed the lowestmean omission error per density and overall meanomission error, whereas the neural networks SOMand fuzzy ARTMap had the highest mean omissionvalues for both. Mean omission errors per classvaried among the algorithms, with the lowest valueTable 2. Omission errors of estimated soil orders using fourmachine learning algorithms and five sampling densities, forthe three neural networks evaluated and for Gini decisiontree.Table 3. Commission errors of estimated soil ordersusing four machine learning algorithms and five samplingdensities, for the three neural networks evaluated and forGini decision tree.Soil orderSoil orderResults and losMean0.540.360.280.560.43Points per hectare1.524Fuzzy g maps .250.790.780.760.740.500.470.450.44Multi-layer perceptron .210.790.740.700.740.500.470.450.43Gini decision losNeossolosMean0.650.280.220.640.45Pesq. agropec. bras., Brasília, v.47, n.9, p.1395-1403, set. 2012Points per hectare1.524Fuzzy g map .250.630.550.500.530.480.440.410.40Multi-layer perceptron .190.510.470.480.330.430.390.400.39Gini decision .560.220.210.590.40

Prediction of soil orders with high‑spatial resolutionfor Argissolos found by the fuzzy ARTMap neuralnetwork; for Cambissolos, by the MLP neural network;for Chernossolos, simultaneously by the MLP neuralnetwork and Gini decision tree; and for Neossolos, bythe Gini decision tree (Table 2).Regarding commission errors, except when allclasses could not be estimated, the lowest values forArgissolos, Cambissolos, and Chernossolos were foundusing the Gini decision tree with sampling densities oftwo, four, and four points per hectare, respectively. ForNeossolos, the lowest commission error was obtainedusing the MLP neural network with four points perhectare. The MLP neural network and Gini decision treealso showed the lowest mean commission errors perdensity and overall mean commission error, whereasthe neural networks SOM and fuzzy ARTMap had thehigher mean values for both. Lowest mean commissionerrors per class for Argissolos, Cambissolos, andChernossolos were found by the Gini decision tree, andfor Neossolos by the MLP neural network (Table 3).Omission and commission errors tended to decreaseas sampling density increased (Tables 2 and 3). Therelationship between predictor variables and classes tobe estimated can be better fitted by making decisionrules more consistent and reducing the confusionbetween classes. As observed by Lippitt et al. (2008),this is particularly important for classes with smallextent, which can be subsampled at lower samplingdensities.In almost all cases, Cambissolos and Chernossoloswere the classes with the lowest omission andcommission errors. This was expected, since theseclasses were more likely to be correctly mapped asthey cover most of the study area, i.e., 41.8 and 34%,respectively. However, this reveals some inadequacyof the random sampling scheme adopted in the presentstudy. According to Pal & Mather (2003), not only thesample size is important for classification algorithms,but also the sampling schema. Schmidt et al. (2008)reported that, for small classes, proportional samplingcan return better results than a random schema. Thismay be the case for Neossolos and Argissolos, whoseextension corresponds to only 9.1 and 15.1% of thestudy area, respectively. At lower sampling densities,the number of random samples within classes, withlow occurrence, may not be sufficient to define theappropriate decision rules (Table 1).1399The lowest overall accuracy was found for the neuralnetwork fuzzy ARTMap with sampling density of 0.5point per hectare, while the highest and identical valuewas obtained for both Gini decision tree with 2 pointsper hectare and for MLP neural network with 4 pointsper hectare (Table 4). In fact, the Gini decision treeand MLP neural network performed similarly for allsampling densities, with overall accuracy even above60% and higher than that for the fuzzy ARTMap andSOM neural networks. However, overall accuracy forthe MLP neural network with 0.5 point per hectare ismisleading because the algorithm completely omittedtwo classes. Since the extent of the predicted classescomprises more than 75% of the study area (Table 1),even with the omission of two classes, the percentage ofcorrectly classified pixels was still high. This indicateshow overall accuracy can lead to misinterpretation ofmap reliability if it is not analyzed together with otherindicators, such as omission and commission errors.In this sense, quantity disagreement and allocationdisagreement provide further information about error,as they decompose the overall disagreement, whichcan be defined as 1 minus the overall accuracy, intwo components related to the proportion and to thespatial allocation of the estimated classes, respectively(Pontius & Millones, 2011). In general, thecontribution of quantity disagreement (Figure 1 A) forthe overall error was smaller than that of the allocationdisagreement (Figure 1 B), except for the fuzzyARTMap neural network. For this algorithm, quantitydisagreement was the major component of error and, inmost cases, it was clearly above the other algorithms.Furthermore, its steep curve (Figure 1 A) indicates thatquantity disagreement is highly sensitive to samplingdensity.Among all algorithms, the MLP neural networkwas the less consistent, showing an unstable,Table 4. Overall accuracy (%) of estimated maps using fourmachine learning algorithms and five sampling densities, forthe three neural networks evaluated and for Gini decisiontree.Machine-learning algorithmFuzzy ARTMapSelf organizing map (SOM)Multi layer perceptron (MLP)Gini decision treeMean0.54258636254.0Points per hectare11.5244553576660646668656867716767716959.3 63.0 65.3 68.5Mean52.663.266.967.2-Pesq. agropec. bras., Brasília, v.47, n.9, p.1395-1403, set. 2012

1400E.C. Sarmento et al.nonlinear response both in quantity disagreementand in allocation disagreement. As the number ofsamples increased, an alternation between quantityand allocation was observed. At times, the algorithmdid not estimate the correct proportion of classesand, at others, the estimated proportion was correctbut many pixels were misallocated. Omission errors(Table 2) and commission errors (Table 3) did notreveal this inconsistence. Overall accuracy (Table 4),instead, suggests a better performance, which showsthe importance of considering these two componentsof error, as proposed by Pontius & Millones (2011),when evaluating classifiers.The SOM neural network and Gini decision treehad similar performance, showing the lowest quantitydisagreement among all algorithms. Their flat curves(Figure 1 A) also indicate a low dependence on thenumber of samples for this component of error.Allocation disagreement, however, was higher thanFigure 1. Quantity (A) and allocation disagreement (B) ofestimated maps using four machine-learning algorithms andfive sampling densities. ARTMap, fuzzy adaptive resonancetheory; MLP, multi-layer perceptron; SOM, self-organizingmap; DT, Gini decision tree.Pesq. agropec. bras., Brasília, v.47, n.9, p.1395-1403, set. 2012that for the fuzzy ARTMap and MLP neural networks,showing a weak response on sampling density(Figure 1 B), with the Gini decision tree presentinglower values. Both the SOM neural network and Ginidecision tree were relatively stable in relation to thenumber of samples and tended to predict classes withthe correct proportion, but misallocated some pixels.Visual analysis showed that part of the misallocationsoccurred close to the boundaries of classes. Accordingto Grimm & Behrens (2010), this is expected becausethe conventional reference map was drawn by hand,whereas algorithms used fixed rules to predict classeson the whole map. As a consequence, some discordanceis common near the boundaries, and, in these cases,prediction may be more reliable than the conventionalmap.Regarding the magnitude for omission andcommission errors and overall accuracy, values weresimilar to those reported by Coelho & Giasson (2010)for decision trees in predicting soil classes from terrainvariables at a coarser spatial resolution. Values foundfor overall accuracy in the present work were higherthan those obtained by Giasson et al. (2011), whenpredicting soil classes with high spatial resolutionfrom terrain variables using several decision trees.Both studies were developed in similar subtropicalconditions, but used a fixed number of samples. Zhaoet al. (2009) obtained overall accuracy above 80%using neural networks to predict sand, clay, and siltcontents with high spatial resolution, whereas the bestvalue found in the present work was 71%, for the MLPneural network and Gini decision tree.In some aspects, these results partially disagree withLippitt et al. (2008), who reported better performancefor the SOM neural network, when compared to MLP,in classifying remote sensing data. In the presentwork, MLP showed higher overall accuracy thanSOM (Table 4). This may be due to the intrinsiccharacteristics of the predictor and estimated dataset,as well as to differences in the configuration of theneural network structure used. However, SOM wasmore consistent in terms of quantity disagreementand allocation disagreement (Figure 1) and, therefore,should be preferred (Lippit et al., 2008). This resultsis in accordance with Srinivasulu & Jain (2006), whorecommend that performance evaluation should be doneusing a wider variety of indicators rather than relyingonly on a few general statistics, usually employed.

Prediction of soil orders with high‑spatial resolutionLippitt et al. (2008) also observed that under optimalsampling, with a high number of samples, differentclassifiers usually show low and close error valuessimultaneously for a specific dataset. In this case,the number of samples probably is near a limit fromwhich increasing sampling density will not add usefulinformation and, moreover, can generate overfitting(Hjort & Marmion, 2008). This may be valid for theGini decision tree, whose overall accuracy decreasedand whose commission errors and quantity disagreementincreased at the highest sampling density. The samplingdensity of four points per hectare matches the upperlimit of the recommended range of field observationsfor detailed soils surveys in Brazil (Manual técnico depedologia, 2007). Therefore, conventional soil surveysampling schemas may be a helpful guide to drive fielddata collection for DSM, at least for detailed scales asused in the present study.This is relevant when thinking of operationalprocedures for DSM, since the sampling strategyis a vital issue for the quality of the training data.The more representative samples are introducedto a classification process, the more accurate andreliable results will be produced (Kavzoglu, 2009).In the present work, an available soil map was usedas reference data, which allowed evaluating thealgorithm's performance in response to the numberof samples aiming future applications; however, inpractice, most samples must be collected on the field.The challenge is to obtain a representative sampleset large enough so that no relevant information getslost, but as sparse as possible in order to save labor,time, and costs. In this case, representative means thatboth size and quality of the sample data are equallyimportant. Therefore, knowledge on the performance,sensitivity, and reliability of classification algorithmsis important to define appropriate sampling (Schmidtet al., 2008; Kavzoglu, 2009).In general, the Gini decision tree was less sensitiveto sampling density than the three neural networksused, and the fuzzy ARTMap neural network showedthe highest sensitivity among all algorithms. Forthe MLP and SOM neural networks, some indicatorswere contradictory. Omission and commission errorsand overall accuracy indicate that the MLP neuralnetwork performed better than SOM, but MLP showeda critical minimum for sampling density below whichit could not estimate all classes. However, quantity1401disagreement and allocation disagreement indicatedthat the SOM neural network was the most consistentamong all used algorithms, whereas MLP was quiteinconsistent. Since the Gini decision tree yields higheraccuracies with lower sampling densities, it seemsto be the most advantageous choice for predictingoccurrence of soil orders, at high spatial resolution,in the study area.Disregarding differences on algorithm performance,all estimated maps showed more spatial details thanthe conventional soil map used as a reference, whichagrees with previous studies (Zhu, 2000; Hempelet al., 2008). This was expected, since conventionalsoil maps are restricted to a minimum mapdelineation size. In Brazil, the minimum mappablearea for detailed soil surveys is set to 1.6 ha (Manualtécnico de pedologia, 2007). However, in the presentstudy, the classification algorithms predicted smallerspatial units, since prediction was done on a pixelwith 5 m of spatial resolution. Once a predictionmodel is fitted using the selected variables, it is thenuniformly applied to the whole area to be mapped.In conventional surveys, unvisited places must beinferred from soil‑landscape relations observed atother locations, which is a less consistent, subjectiveprocess. Therefore, in many cases, predicted classesmay be

soil properties or the occurrence of soil classes in a reliable way, broadly referred to as digital soil mapping (DSM). According to Lagacherie (2008), DSM can be defined as the creation and population of spatial soil information systems using numerical models that infer spatial and temporal variations of soil properties and