An Assessment of a Random Forest Classifier for a Crop Classification Using Airborne Hyperspectral Imagery

Abstract

Crop type classification is essential for supporting agricultural decisions and resource monitoring. Remote sensing techniques, especially using hyperspectral imagery, have been effective in agricultural applications. Hyperspectral imagery acquires contiguous and narrow spectral bands in a wide range. However, large dimensionality results in unreliable estimates of classifiers and high computational burdens. Therefore, reducing the dimensionality of hyperspectral imagery is necessary. In this study, the Random Forest (RF) classifier was utilized for dimensionality reduction as well as classification purpose. RF is an ensemble-learning algorithm created based on the Classification and Regression Tree (CART), which has gained attention due to its high classification accuracy and fast processing speed. The RF performance for crop classification with airborne hyperspectral imagery was assessed. The study area was the cultivated area in Chogye-myeon, Habcheon-gun, Gyeongsangnam-do, South Korea, where the main crops are garlic, onion, and wheat. Parameter optimization was conducted to maximize the classification accuracy. Then, the dimensionality reduction was conducted based on RF variable importance. The result shows that using the selected bands presents an excellent classification accuracy without using whole datasets. Moreover, a majority of selected bands are concentrated on visible (VIS) region, especially region related to chlorophyll content. Therefore, it can be inferred that the phenological status after the mature stage influences red-edge spectral reflectance.

1. Introduction

Crop type classification is crucial for supporting agricultural decisions and monitoring, e.g. crop area and yield estimation (Löw et al., 2013). Moreover, it is important for food security, so crop mapping is increasingly in demand for society (Hao et al., 2015). The traditional approach for crop mapping is based on enumeration by field surveys. However, improvements in crop productivity, using surveys is challenging due to their intensive and time-consuming labor requirements. Therefore, remote sensing technology can be an effective tool for evaluating agriculture, and has developed into precision agriculture. Precision agriculture can handle extensive data and process in a timely manner based on the interaction of electromagnetic radiation with crop or soil (Mulla, 2013). In particular, hyperspectral imagery has been introduced to precision agriculture due to its acquisition in contiguous and narrow spectral bands(Chutia et al., 2016). Moreover, it collects the spectral response of crop or soil in a wide spectral range, which would not possible with multispectral imaging (Thenkabail et al., 2004). These characteristics can improve in-depth understanding of specific crop characteristics and discriminate crop types(Sahoo et al., 2015). However, too many spectral bands cause several difficulties for processing hyperspectral data.

The main difficulty is called the “Hughes phenomenon,” & “curse of dimensionality.∓rdqu; With the increasing dimensionality of hyperspectral data, the number of required training samples per class must be exponentially increased as well to obtain reliable and accurate estimate statistical properties for a classifier (Hsu, 2007). To overcome this limitation, there are two categories for reducing the dimensionality of hyperspectral data: feature extraction (based on transforming original dataset into a reduced number of new features which contain high proportion of original information) and feature selection (based on selecting a subset from original dataset which maintains useful information) (Pal and Foody, 2010). By reducing the dimensionality of hyperspectral data, it is possible to alleviate the computational burden and deepen our understanding of the most suitable spectral bands for crop mapping (Chan and Paelinckx, 2008).

The Random Forest (RF) classifier is an ensemble-learning algorithm, which is based on Classification and Regression Tree (CART). The RF has received increasing attention due to the excellent classification results and fast processing speed over the last two decades (Belgiu and Drăguţ, 2016). In particular, a number of studies have investigated excellent RF performance with respect to high accuracy in crop classification (Pal, 2005; Ok et al., 2012; Sonobe et al., 2014). Moreover, RF has been applied to hyperspectral data due to its advantage of relatively insensitive to the small sample size and low computation burden (Larence etal., 2006; Belgiu and Drăguţ, 2016). Also, RF can be adopted for the feature selection method in addition to classification purposes by measuring an explanatory power of each feature. The major advantage is the data-driven knowledge discovery of important bands (Archibald and Fann, 2007), so it is possible to analyze and interpret a classification model. However, only a few studies have applied RF for feature selection while managing hyperspectral data (Chan and Paelinckx, 2008; Dye et al., 2011; Adam et al., 2012). Therefore, selection of hyperspectral bands by applying RF for feature selection is needed to assess crop classification.

The objective of this study was to evaluate RF performance in a crop classification using airborne hyperspectral imagery by determining the best hyperspectral bands to discriminate crops, assessing classification accuracies using selected bands, and studying the role of selected bands related to the phenological status on the flight date. In this study, “variable” and “feature” refer to the “spectral band.” The remaining sections of this paper are organized as follows. Section 2 introduces a description of the study area and data followed by method of this study. Results with respect to the effects of parameter optimization, variable importance, and accuracy assessment are introduced in Section 3. Finally, conclusions and a description of future works are presented in Section 4.

2. Mthodology

1) Study area and data

The experimental study area was the cultivated areas of Chogye-myeon, Habcheon-gun, Gyeongsangnamdo, South Korea. Airborne hyperspectral imagery collected using the AISA FENIX sensor was acquired on 25 May, 2017. Field survey for acquiring reference data was conducted at the same time of image data collection. The AISA FENIX sensor has a spatial resolution of 1.5 m and collects in 450 bands ranging from 380 to 2500 nm. The spectral range includes Ultraviolet (UV, 380-400 nm), Visible (VIS, 400-700 nm), Near infrared (NIR, 700-1300 nm), and Shortwave infrared (SWIR, 1300-2500 nm). The main characteristics of hyperspectral dataset are presented in Table 1.

Table 1. Characteristics of hyperspectral dataset

The dominant crops grown in the area are garlic, onion, and wheat. Garlic and onion are two ofthe most important spicy vegetablesin Korea, while wheat is one of the important food crops among the world with rice and corn. According to Rural Development Administration (RDA, 2007), plant to harvest stage of each crop are September to June (garlic, onion) and Octoberto June (wheat). However, by investigation on the field survey, garlic and onion were in the harvest stage, and wheat was in the mature stage on the flight date. There were also bare soil and water afterthe garlic and onions had harvested.In this study, the experiment was conducted on the crop area for the purpose of discrimiating crops. Fig. 1 shows te location of study area.

Fig. 1. Location of study area. Left and Middle show cultivated areas of Chogye-myeon, Habcheon-gun, Gyeongsangnam-do, South Korea where airborne hyperspectral imagery was collected. Right shows the test site of this study where reference data were collected by field survey.

2) Method

Fig. 2 shows the flowchart of this study. At first, the raw data was radiometrically and geometrically corrected using CaliGeoPRO 2.2 software, and atmospherically corrected using ATCOR-4 (Atmospheric/Topographic Correction for Airborne Imagery) module. Also, noisy bands (1-7 and 440-450 band) and water vapor bands (244-275 and 320-361 band) were eliminated manually. Then, subset of test site was chosen where field survey was conducted. Image size of subset data is 560×634 (pixels). After preprocessing and subset step, RF parameter optimization was then conducted, followed by selecting 25 and 50 highest ranking bands based on variable importance of RF. In this study, reference data were divided randomly into training (5%) and validation (95%) sets. The training data were used to train the RF classifier and the validation data were used for the accuracy assessment. Finally, accuracy assessment was conducted on all, 25, and 50 highest ranking bands. Following sections present more details about each step.

Fig. 2. Flowchart.

(1) Random Forest

The RF is anensemble learning algorithm basedon CART, which was proposed by Breiman (Breiman, 2001). Each tree is built based on random bootstrapped samples of training data, called a bagging approach. About two-thirds of the samples, called the in-bag samples, are used to build the trees. The remaining onethird of the samples, called the out-of-bag (OOB) samples, are put down the trees to estimate the misclassification error, called the OOB error. This error evaluates the internal performance of the RF model.

Each node is split using random subsets of total variables. The variable that yields the maximum reduction in impurity is selected to split the samples at each node among the random selected variables.In this study, the Gini Index was used as a criterion to perform splitting and can be written as the equation (1):

\(\sum \sum_{j \neq i}\left(\frac{f\left(C_{i}, T\right)}{|T|}\right)\left(\frac{f\left(C_{j}, T\right)}{|T|}\right)\)       (1)

where is \(\frac{f\left(C_{i}, T\right)}{|T|}\) the probability that training set T belongs to class C_i. The Gini Index measures the class homogeneity and splitting isterminated when the Gini Index is zero (Ok et al., 2012). After the RF model is built, the final classification result is determined by a majority vote among the trees.

The RF classifier requires two user-defined parameter, N (tree) and M (try). N (tree) is the number of decision trees to be generated and M (try) is the number of variables to be selected at each node. I is essential to optimize the parameters to maximize model accuracy (Tatsumi et al., 2015). In this study, a gridsearch approach based on the OOB error was used to find the optimal combination. The combination with the lowest OOB error was selected for the optimal parameters. Based on the previous studies, N (tree) was varied from 100 to 500 with an interval of 100, and M (try) was set to {1/3, 1/2, 1, 2, 3}*(default value) (Adam et al., 2012; Akar and Güngör, 2013; AbdelRahman et al., 2014; Adam et al., 2014). Default value of M (try) is the square root of the number of input variables. In this study, RF classification, including its parameter optimization process, were run on the Interactive Data Language (IDL)-based ImageRF in the EnMAP-Box (Waske et al., 2012).

(2) Variable Importance of RF

During the classification process, OOB samples were used to determine relative importance of the variables. To evaluate the importance of each variable, RF calculated the OOB error after permuting one of the random variables while the other variables were maintained. Differences in OOB error after permuting a randomly selected variable were used as measures of importance. Larger differences in OOB error indicated that the permuted variable was the more important variable. RF model was built ten times, and varible importance was averaged over the 10 models due to the randomness in the variable importance statistics (Löw et al., 2013). In this study, the 25and 50 highest ranking bands were selected.

(3) Accuracy assessment

The classification results were evaluated in terms of overall accuracy (OA), kappa coefficient, producer’s accuracy (PA), and user’s accuracy (UA) by calculating the confusion matrix. OA indicates the total proportion of correctly classified samples. Kappa coefficient is a measure of classification performance considering the chance agreement between the prediction and reference data, and can be written as equation (2):

\(\kappa=\frac{p r_{o}-p r_{e}}{1-p r_{e}}\)       (2)

where is pr_o the probability of correct classification, and pr_e is the probability of chance agreement. A perfect agreement with no chance agreement would result in κ=1. PA indicates the proportion of a certain reference class being correctly classified, while UA indicates the probability that a certain labelled sample being correctly classified.

3. Experimental Results

1) Effects of RF parameters on OOB error

Before performing variable selection, it is necessary to optimize the parameters, N (tree) and M (try), based on the grid search method using the OOB error. Table 2 shows that the lowest OOB error is 0.024%, but there are multiple combinations. Finally, N (tree) and M (try) values of 100 and 6 are chosen after comparing the OA and kappa coefficient.

Table 2. Parameter optimiztion results

2) Variable selection using the OOB method

After optimizing parameters, the 25 and 50 highest ranking bands were selected using the OOB method. Fig. 3 shows the mean decrease in OOB error when each variable is permuted on the OOB set, which is the importance of each variable to the RF classification.

Fig. 3. Variable Importance of RF.

Table 3 showsthe 25 and 50 highest ranking bands and their wavelength center. Among the selected bands, the 25 highest ranking bands consist of nineteen from VIS and six from NIR, and the 50 highest ranking bands consist of thirty-eight from VIS, eleven from NIR, and one from SWIR. A majority of selected bands are concentrated on VIS regions and only one SWIR band is included in the 50 highest ranking bands. Also, six in the 25 highest ranking bands and nine in 50 highest ranking bands are related to the red-edge region (700-750 nm) which is sensitive to vegetation stress and dynamics. Moreover, nine in the 25 highest ranking bands and sixteen in the 50 highest ranking bands are included in the region (410-430 and 600-690 nm) where Chlorophyll a presents maximum absorption (Sahoo et al., 2015).

Table 3. 25 and 50 highest ranking bands and wavelength center (unit : nm)

3) Accuracy Assessment

After reducing the dimensionality of the hyperspectral data, an accuracy assessment was performed by calculating the confusion matrix. Table 4, 5, 6 show tht the OA of the RF classifier using all bands is 92.51% with a kappa coefficient of 0.8853, using the 25 highest ranking bands is 92.63% with a kappa coefficient of 0.8871, and using the 50 highest ranking bands is 93.03% with a kappa coefficient of 0.8931. Among three cases, using 25 and 50 highest ranking bands has larger classification accuracy than using whole datasets. Misclassification appears apparent between onion and garlic, and Fig. 4 shows that misclassification between onion and garlic is apparent.

Table 4. Confusion Matrix of RF using all bands

Table 5. Confusion Matrix of RF using 25 highest ranking bands

Table 6. Confusion Matrix of RF using 50 highest ranking bands

Fig. 4. Classification maps obtained using RF classifier (a) Ground truth, (b) All bands, (c) 25 highest ranking bands, (d) 50 highest ranking bands.

Table 7 shows that the difference between UA and PA for the RF classifier using the 25 highest ranking bands is relatively larger than the difference using all bands, and using the 50 highest ranking bands has the largest difference among the three cases. Moreover, UA and PA for onion and garlic are relatively lower than those for wheat. In additin to contributing to a misclassification between onion and garlic, the lower UA and PA are due to the sparse distribution of onion and garlic on the soil, resulted in high intra-class variability.

Table 7. UA and PA of the studied classes using all, 25, 50 highest ranking bands

In this study, RF was performed as a classification method and variable selection. Before applying RF classification, parameters, N (tree) and M (try), were tuned based on the combination that resulted in the lowest OOB error. Moreover, variables were selected based on estimating variable importance using a permutation of variables. Experimental results show that the RF has benefit for selecting important bands on crop classification. Compared to full datasets with OA of 92.51% and kappa coefficient of 0.8853, OA for RF classifier using the 25 highest ranking bands is 92.63% with a kappa coefficient of 0.8871, and OA for RF classifier using the 50 highest ranking bands is 93.03% with a kappa coefficient of 0.8931. Therefore, using the selected bands shows an excellent classification accuracy without use oflarge numbers of whole bands.

In the previous studies, the utility of red edge and SWIR region has significantly increased for crop classification due to its high correlation with leaf chlorophyll and water content respectively (Thenkabail et al., 2004; Chan and Paelinckx, 2008; Kim and Yeom, 2014). In this study, the importance of SWIR region has not shown sinificant, but red edge region and region related to chlorophyll a have shown significant for discriminating crops. The reason is not clear, but it is possible that red reflectance of wheat increases due to senescing veetation while amounts ofleaf chlorophyll and water content ofthe other crops decrease (Thenkabail and Lyon, 2016). Also, it can be inferred that the phenological status after the mature stage influences red-edge spectral reflectance rather than SWIR.

Moreover, mono-temporal classification accuracy is strongly related to the phenological status of the crops and weather conditions of the acquisition date. Misclassification between garlic and onion is relatively high due to their similar crop calendars. Therefore, additional research is needed to investigate multitemporal datasets including Vegetation Indices (VIs) or phenological metrics for increasing classification accuracy and better understanding the influential factor on each growing season.

4. Conclusion

Crop type identification using the RF classifier with airborne hyperspectral imagery was performed in this study. Hyperspectral data has contiguous and narrow spectral bands in a wide range, from UV to SWIR regions, which has the potential for discriminating agricultural crops with high accuracy. However, large dimensionality results in unreliable estimates from a classifier and high computational burdens. Therefore, reducing dimensionality of hyperspectral imagery is essential. Inthis study, RF was used to reduce data dimensionality. RF is well known for its high classification accuracy, fast processing speed and approach for feature selection. Therefore, as well as a feature selection method, RF wa adopted to classify data using all bands, the 25 highest ranking bands, and 50 highest ranking bands. As a result,RF classification using the 25 and 50 highest ranking bands result in better accuracy than using all bands. Moreover, a majority of selected bands are concentrated on VIS regions and large portion of spectral bands are related to the chlorophyll content. Therefore, it can be inferred that the phenological status after the mature stage influences red-edge spectral reflectance. In the future work, multi-temporal datasets including VIs or phenological metrics are needed to investigate for increasing classification accuracy andmore information about the influential factor on each growing season.