Testing the spectral resolutions of the new multispectral sensors for detecting Phaeosphaeria leaf spot (PLS) infestations in maize crop

Maize is one of the most important subsistence and commercial crops in the world. In Africa, it is regarded as one of the most popular food crops. Recently however, significant losses due to Phaeosphaeria leaf spot (PLS) infestation have been reported. Therefore, techniques for early detection of PLS infestation are valuable for mitigating maize yield losses. Recently, remotely sensed datasets have become valuable in crop assessment. In this study, we sought to detect early PLS infestation by comparing the performance of commonly used higher spatial resolution sensors (WorldView, Quickbird, Sentinel series 2, RapidEye and SPOT 6) based on their spectrally resampled field spectra. Canopy training spectra were collected on leaves with signs of early infestation and healthy leaves spectral characteristics used for comparison. Training data was collected in 2013 growing season while test data was collected under similar conditions in 2014. The Random Forest algorithm was used to establish the Kappa and overall, user and producer's accuracies. Results showed that the RapidEye sensor with an overall classification accuracy of 86.96% and Kappa value of 0.76 performed better than the rest of the sensors while the Red, Yellow and Red-Edge bands were most useful for detecting early PLS infestation. The value of the RapidEye sensor in detecting early PLS infestation can be attributed to the optimally centred Red Red-Edge bands sensitive to changes in chlorophyll content, a consequent of PLS infestation on maize leaves. The study provides valuable insight on the value of existing sensors, based on their sensor characteristics in detecting early PLS infestation.


Introduction
Globally, maize is regarded as one of the most important subsistence and commercial crops. In Africa, a number of authors (eg Derera et al. 2007, Sibiya et al. 2011and Benhin 2008 note that maize remains the most important food crop. However, recently maize production in the tropical and sub-tropical regions of the continent has significantly declined due to the Phaeosphaeria leaf spot (PLS) disease infestation (Gonçalves et al., 2013;Moreira et al. 2009). The disease affects maize foliar and is caused by the ascomycete fungus Phaeospharia maydis (Henn.). Initially spotted in India, studies show that it has spread to North and South America, and recently east and southern Africa (Gonçalves et al., 2013;Moreira et al., 2009;Carson, 2005;Sibiya et al., 2011;Derera et al., 2007).
The infestation often leads to premature leaf desiccation and consequent reduction in the crop cycle, decrease in grain volume and mass and under severe conditions death of the whole plant (Gonçalves et al., 2013;Paccola et al., 2001). Previous studies have shown that substantial amount (11-60%) of grain yield is lost in cultivars sensitive to PLS (Derera et al., 2007;Paccola et al., 2001). Whereas losses due to PLS have not been determined in South Africa, agronomic conditions and preferred varieties suggest vulnerability to extensive infestation and possible significant yield losses.
Consequently, early detection of infestation and adoption of appropriate mitigation measures is necessary for sustaining subsistence and commercial production.
Traditionally, field visual surveys have been used to determine PLS infestations. This method requires constant monitoring and is often costly, time consuming and impractical on extensive fields (Al-Hiary et al., 2011;Liu, et al., 2008). However, there has been an increased awareness on the value of remotely sensed datasets, particularly crop leaf spectral reflectance characteristics for agricultural applications (Chen et al., 2010;Abdel-Rahman and Ahmed, 2008). Recent advances in sensors on satellite, aerial and ground based platforms in concert with robust algorithms and analysis techniques have further made it possible to reliably determine crop bio-physical status, due to among others change in agronomic conditions and pest and disease infestation (Upadhyay et al., 2012;Pinter et al., 2003). The adoption of remotely sensed datasets in disease mapping has particularly attracted a lot of interest (Estep et al., 2004;Jackson et al., 1986;Baret et al., 2007). Such applications are premised on the fact that disease infestation cause physiological alteration on leaves, hence a change in spectral reflectance (Estep et al., 2004;Jackson et al., 1986;Baret et al., 2007).
Whereas there has been a recent proliferation of "new generation" multispectral sensors (e.g. WorldView-2, Sentinel series 2 and RapidEye) with higher spectral resolution in addition to traditional sensors (e.g. Quickbird and SPOT), choosing the most suitable sensor and bands for early detection of PLS infestation remains a challenge. This is because existing sensors are characterised by a different number of bands and unique spectral configurations, which are sensitive to different vegetation properties. Therefore, in this study, we sought to determine the potential of commonly used and proposed sensors for detecting early PLS infestation. Specifically, we sought to compare the performance of the commonly used multispectral sensors and the importance of each of their respective bands in detecting early PLS infestation by resampling hyperspectral field data.

Study Area and field spectral data collection
This study was conducted at the KwaZulu-Natal Department of Agriculture -Cedara Experimental Farm (29°32 S and 30°16 E), in Pietermaritzburg, KwaZulu-Natal province, South Africa ( Figure 1).

Figure1. The generalised study area showing the Cedara experimental farm.
Field spectral data was collected on 5 th of January 2013 and 10 th of January 2014 from 138 plots measuring 3m, 0.75m and 0.3m and 0.75m intra row spacing and row length, respectively at full canopy cover. Plots for the study were chosen based on the quantity of leaves with signs of early PLS infestation using optical evaluation (Sibiya et al., 2011). Plots with less than 10% of the leaves with PLS were regarded as healthy while plots with between 80 to 100% of the leaves with signs of PLS were regarded to be in the early stage of infestation.
Spectral reflectance for early stage PLS maize canopies were acquired using an Analytical Spectral Device (ASD Inc, Boulder, CO, USA) FieldSpec®3spectrometer. Spectral data was gathered under sunny and cloud-free conditions between 10h00 am and 14h00 pm local time in the two growing seasons (5 th of January 2013 and 10 th of January 2014). The ASD spectral measurement range is 350-2500 nm, measuring radiation at 1.4 nm intervals for 350-1000 nm and 2.0 nm for 1000-2500 nm spectral regions. Fifteen to twenty measurements were taken from every plot. The Spectral measurements were then averaged to represent the ultimate spectral measurements of each plot. In every 10 to 20 computations, a white based spectral measure was executed on the calibration panel to ameliorate any change in atmospheric conditions and sun irradiance. In total, 66 plots of early infestation and 72 plots of the healthy maize crop were represented separately. Figure 2 provides a visual depiction of healthy leaves, early PLS infestation and respective spectral characteristics. On the 10 th of January 2014 (the following growing season) field spectral measurements were acquired in the same location using a similar procedure and comparative conditions. Spectral measurements of 60 and 62 plots for early infestation and healthy maize crop were acquired separately.

Field spectra to sensor resampling
Spectral reflectance data was resampled to SPOT6, RapidEye, Quickbird, WorldView-2 and Sentinel-2 to band centres using ENVI 4.7 image processing software (Table1). The method used a Gaussian model with a full width at half maximum (FWMAP) equal to the band spacing provided (Mutanga and Skidmore 2005). Spectral reflectance was resampled to the popular multispectral sensors to determine if their respective spectral bands can be used to map the early PLS infestation.
If the results are positive, the mapping and monitoring of PLS could then be operational at lower cost using these sensor platforms compared by the use of hyperspectral data.

Statistical analysis
Random Forest (RF) algorithm was used to classify and predict healthy maize and the early PLS infestation. Random Forest is a bagging process in which various classification trees are created with reference to random subsets of samples arising from training data (Adelabu et al., 2013). The algorithm classifies and measures important variables in high dimensional data such as hyperspectral data and therefore decreases the "curse of dimensionality" without losing important information in the dataset (Adam et al., 2012). For an exhaustive explanation of the Random Forest algorithm see Breiman (2001). In this study, the RF classifications were computed using the 'rattle' package available in R software and used to generate confusion matrices that provide overall accuracy (OA), user accuracy (UA) and producer accuracy (PA). Cohen's Kappa coefficient were calculated using the formula: Where Pr(a) is the actual detected agreement and Pr(e) represents unpredicted agreement (Hallgren, 2012). Kappa gives measured values of agreement between varying observations (Viera and Garrett, 2005). It provides a determination of variability that range from -1 to +1 where +1 represents absolute agreement between different observations, 0 representing an agreement that can happen by coincidence and -1 represent lower possibility of agreement between observations (Viera and Garrett, 2005).

Determination of variable importance
Variable importance in large datasets identifies a variable that yields outstanding performance compared to other variables. Variable importance in RF computes different measures such as how often a variable is determined, the Gini significance and permutation significance. The permutation importance is considered better than other measures because it assesses value of the variable using mean decrease accuracy based on out-of-bag postulations (Breiman, 2001 Table 2).

Accuracy Assessment
Table2. Accuracy Assessment for early stage Phaeosphaeria leaf spot (PLS) infestation using different sensors. User Accuracy (UA), Producer Accuracy (PA), Overall Accuracy (OA). Healthy  Figure 3 shows the performance of sensors rated in terms of variable importance of their respective spectral bands. Based on the mean decrease percentages, the Red band was identified as the most significant in detecting early PLS infestation. Other significant bands include the Yellow and the Red-Edge bands.

Discussion
This study sought to discriminate between early PLS infestation and the healthy maize crop by resampling hyperspectral data to SPOT 6, Quickbird, RapidEye, Sentinel-2 and WorldView-2.
Spectral resampling was performed using the Random Forest, a robust algorithm for predicting and classifying variables. Results in this study show that RapidEye, with an overall accuracy of 86.96% and Kappa of 0.76, offer better detection of early PLS infestation than other sensors. These findings are consistent with Adam et al. (2012) who found that using relevant bands without redundancy increases classification accuracy in vegetation mapping. According to Adam et al. (2012) Zhang et al., 2005 andApan et al., 2005) have noted the value of the Red band in vegetation studies. Zhang et al. (2005) for instance noted that the Red band was sensitive to crops infested by diseases while Apan et al. (2005) found that the Red band in the visible region was valuable in predicting Alternaria solani fungal infection on tomatoes. The Yellow band's ability to determine early PLS infestation can be attributed to its ability to identify "yellowness'', a common characteristic in diseased plants (Digital Globe, 2010;Oliveira et al., 2010 ). The Red-Edge band has the ability to assess plant nutrition, health and discriminate vegetation species (Filella and Peñuelas, 1994;Pinar and Curran, 1996;Daughtry et al., 2000;Rodriguez et al., 2006). According to Eitel et al. (2011), the Red-Edge band can determine both the amount of chlorophyll and Nitrogen, the main determinants of the crop health.
Results obtained from this study demonstrate the potential of multispectral resolution imagery, such as RapidEye in early detection of PLS infestation. The WorldView-2 is also recommended for PLS detection for its three important bands (Red, Yellow and Red-Edge). This study provides an indication of the most useful sensors for PLS infestation. However, for image applications, this must results should be treated with caution as spatial resolution is also known to significantly influence mapping accuracy. The influence of spatial resolution on the optimum sensor's (eg RapidEy) mapping accuracy needs further investigation.

Conclusion
The aim of this study was to detect early PLS infestation by resampling spectral data to different multispectral sensors. Results in this study showed that early PLS infestation can be better detected using RapidEye's multispectral bands. The significant bands for early PLS infestation were the Red, Yellow, and the Red-Edge bands. Using the Random Forest algorithm, an overall accuracy of 86.96%