*2.3. UAV Multi-Spectral Imagery Acquisition*

The surveys were done using a DJI Phantom 4 quadcopter (DJI Innovations, Shenzhen, China). This UAV was equipped with a five-band multispectral camera MicaSense RedEdge MTM (MicaSense, Inc., Seattle, WA, USA), which has five narrow bands: Blue (465–485 nm), green (550–570 nm), red (653–673 nm), red edge (712–722 nm), and near-infrared (800–880 nm). The flight at Guangxi site was conducted between 12:30 p.m.–13:30 p.m. on 7 August 2018 and covered an area of 21 ha. The flight at Hainan site was conducted between 11:00 a.m.–12:00 p.m. on 11 December 2018 and covered an area of 11 ha. Both the flight plans were developed to ensure greater than 80% cross-track and along-track overlap rates. The multispectral imagery was acquired from a flying height of 120 m above the ground with a ground sample distance of 0.08 m. Pre- and post-flight images of MicaSense calibrated reflectance panels with known reflectance were also captured using the RedEdge sensor to aid in radiometric conversion.

### *2.4. Data Analysis*

In this study, different VIs were used to identify the infestation status of Fusarium wilt in banana. BLR was used to assess the spatial relationship between the VIs and the plants infested or not infested with Fusarium wilt. In order to assess the classification accuracy of images with different spatial resolutions, we chose to resample the original UAV imagery of 0.08 m to generate images with 0.5-m, 1-m, 2-m, 5-m, and 10-m resolutions using the nearest neighbor resampling algorithm. These resolutions were selected because they were similar to those of several mainstream and easily accessible satellite imagery products (i.e., WorldView series with a resolution of 0.5 m, GF-2 with a resolution of 1 m, GF-1 and GF-6 with a resolution of 2 m, RapidEye with a resolution of 5 m, and Sentinel-2 with a resolution of 10 m) for agricultural applications.

#### 2.4.1. Vegetation Indices

Considering the potential pathological characteristics of the Fusarium wilt disease infestations, eight VIs related to pigment absorption and plant growth were selected to characterize the biochemical and biophysical variations caused by individual infestations. The VIs included the NDVI, normalized difference red edge index (NDRE), green chlorophyll index (CIgreen), red-edge chlorophyll index (CIRE), structural independent pigment index (SIPI), red-edge structural independent pigment index (SIPIRE), carotenoid index (CARI), and anthocyanin reflectance index (ARI). The definitions of these VIs are listed in Table 1.


**Table 1.** List of eight vegetation indices (Vis) used in this study.

#### 2.4.2. Binary Logistic Regression

BLR is one of the most frequently used multivariate analysis methods, where the dependent variable is a binary variable representing the presence or absence of an event. The dependent variable in the BLR method is a function of the probability and is expressed as [32]:

$$p = 1/\left(1 + e^{-y}\right) \tag{1}$$

where *p* is the probability of disease occurrence, which ranges from 0 to 1, *y* is the linear combination, and *e* is the numerical constant. The *y* can be expressed by formula as:

$$y = \beta\_0 + \beta\_1 \mathbf{x}\_1 + \beta\_2 \mathbf{x}\_2 + \dots + \beta\_n \mathbf{x}\_n \tag{2}$$

where β<sup>0</sup> is the intercept and β*<sup>i</sup>* and *x<sup>i</sup>* (*i* = 0, 1, 2,..., *n*) are the slope coefficients independent variables, respectively. In this study, the BLR method was used to establish the spatial relationships between the plants infested or not infested with Fusarium wilt and the VIs extracted from different resolution images. The modelling dataset were used to fit the logistic regression models through SPSS 20.0 software (SPSS Inc., Chicago, Illinois, USA).

#### 2.4.3. Accuracy Assessment

After the model fitting, two validation datasets (VD1 and VD2) were used to quantitatively evaluate the disease identification accuracy, respectively, with indicators such as the overall accuracy (OA) and the Kappa coefficient [43,44]. The OA is the sum of the correctly classified plots divided by the total number of plots. The Kappa value ranges between −1 and 1 with a larger value indicating better model performance. Model performance can be judged as excellent if kappa ≥ 0.75, good if 0.75 > kappa ≥ 0.4, or poor if kappa < 0.4 [45].

#### **3. Results**

#### *3.1. Statistical Characteristics of Samples*

We analyzed the differences in the VI values between the healthy and diseased samples obtained from the Guangxi site and Hainan site, and conducted independent *t*-test analyses for each sample. Table 2 shows the statistical characteristics of the VI values of the healthy and diseased samples. The results showed that there were significant differences in the values of CARI, CIgreen, CIRE, NDVI, NDRE, and ARI between the healthy and diseased samples (*p* < 0.01), but no significant differences in the SIPI and SIPIRE values (*p* > 0.05). Therefore, CIgreen, CIRE, NDVI, NDRE, CARI, and ARI were selected for the subsequent analysis.

#### *3.2. Model Fitting with Di*ff*erent Vegetation Indices*

In this study, the modeling dataset was used to fit the logistic regression models describing the relationship between the VIs and the plants infested or not infested with Fusarium wilt. Both the validation dataset 1 (VD1) from the Guangxi site and validation dataset 2 (VD2) from the Hainan site were used to verify the classification accuracy of the fitted models. The results showed that the use of the CIgreen, CIRE, NDVI, and NDRE resulted in relatively good fitting models with an OA greater than 80% (Table 3). Of all VIs, the CIRE obtained the highest validation OA and highest validation Kappa coefficient both for VD1 (OA = 91.7%, Kappa = 0.83) and VD2 (OA = 80.0%, Kappa = 0.59), indicating that CIRE had the best performance for Fusarium wilt identification. For the same type of VI, higher validation OA and Kappa coefficient were obtained for VIs that included the red-edge band (e.g., CIRE vs. CIgreen, and NDRE vs. NDVI). However, the validation OA and Kappa coefficients based on the CARI and ARI were relatively low.


**Table 2.** Statistical characteristics of VIs values of the healthy and diseased samples.

**Table 3.** The logistic regression models for different vegetation indices.


\* Overall accuracy.

#### *3.3. Model Fitting for Di*ff*erent Resolution Imagery*

The effect of resolution on the identification accuracy of banana Fusarium wilt disease was assessed to provide reference information for large-scale applications of satellite-based data. The UAV imagery was resampled to represent five resolutions (0.5 m, 1 m, 2 m, 5 m, and 10 m) to monitor the occurrence

of Fusarium wilt. In order to consider satellite imagery with red-edge bands, both the optimal VI with a red-edge band (CIRE) and the optimal VI without a red-edge band (CIgreen) were calculated for the images with different spatial resolutions. Table 4 lists the results of logistic regression fitting between locations of infested or noninfested plants and the optimal VIs (CIRE and CIgreen) at different resolutions. The results showed that the logistic regression models for the CIRE for the 0.5-m, 1-m, and 2-m resolution imagery had an acceptable fitting accuracy with the fitting OA greater than 80% (OA = 90.5%, 83.2% and 81.1% for 0.5-m, 1-m, and 2-m resolution, respectively). Verification results also showed that the CIRE for the 0.5-m, 1-m, and 2-m resolution obtained the acceptable validation OA (over 70%) and Kappa coefficient (over 0.40). For the VD1, the validation OA for the 0.5-m, 1-m, and 2-m resolution were 91.7%, 79.2%, and 75.0%, respectively, and the Kappa coefficients were 0.83, 0.60, and 0.53, respectively. For the VD2, the validation OA for the 0.5-m, 1-m, and 2-m resolution were 85.7%, 74.3%, and 71.4%, respectively, and the Kappa coefficients were 0.71, 0.48, and 0.41, respectively. However, the OA and Kappa coefficient for the 5-m and 10-m resolutions were relatively low. As the resolution decreased, the OA and Kappa coefficient showed a decreasing trend. Moreover, at the same spatial resolution, the CIgreen resulted in lower accuracy of the identification models of Fusarium wilt than the CIRE. For the CIgreen, the result was only acceptable at 0.5-m resolution.


**Table 4.** The logistic regression models for the CIRE and CIgreen VIs for images with different resolutions.

\* Overall accuracy.

### *3.4. Mapping Disease Distribution using Imagery with Di*ff*erent Resolutions*

In order to further understand the visual effect of resolution, the distributions of banana Fusarium wilt infested or non-infested regions at Guangxi site were mapped at different resolutions (including 0.5-m, 1-m, 2-m, 5-m, and 10-m resolutions). CIRE and CIgreen were used as input variables to create disease distribution maps based on their identification models of banana Fusarium (Figures 2 and 3). The maps with 0.08-m, 0.5-m, 1-m and 2-m resolutions appeared quite similar with regard to the occurrence of Fusarium wilt disease (Figures 2a–d and 3a–d), whereas the maps with 5-m and 10-m resolutions showed little detail (Figure 2e,f and Figure 3e,f). We also calculated the area and percentage of the Fusarium wilt infected regions based on different resolution maps (see Table 5). For the CIRE-based maps, the areas of Fusarium wilt disease regions were in the range of 5.69–6.59 ha, accounting for 38.2%–44.3% of the total planting area of bananas at different resolutions. Within the 2-m resolution, the percentage of the Fusarium wilt infected regions were in the range of 40.8%–43.6%. For the CIgreen-based maps, the areas of Fusarium wilt disease regions were in the range of 5.09–6.63 ha, accounting for 34.2%–44.6% of the total planting area of bananas. At 0.08-m and 0.5-m resolutions, the percentage of the Fusarium wilt infected regions were 40.1% and 44.6%, respectively.

**Figure 2.** Maps of the distribution of banana Fusarium wilt at Guangxi site based on the CIRE for different resolutions. (a) 0.08-m resolution; (b) 0.5-m resolution; (c) 1-m resolution; (d) 2-m resolution; **Figure 2.** Maps of the distribution of banana Fusarium wilt at Guangxi site based on the CIRE for different resolutions. (**a**) 0.08-m resolution; (**b**) 0.5-m resolution; (**c**) 1-m resolution; (**d**) 2-m resolution; (**e**) 5-m resolution; (**f**) 10-m resolution.


**Table 5.** Area and percentage of the Fusarium wilt infected regions based on different resolution maps.

(e) 5-m resolution; (f) 10-m resolution.

**Figure 3.** Maps of the distribution of banana Fusarium wilt at Guangxi site based on the CIgreen for different resolutions. (a) 0.08-m resolution; (b) 0.5-m resolution; (c) 1-m resolution; (d) 2-m resolution; **Figure 3.** Maps of the distribution of banana Fusarium wilt at Guangxi site based on the CIgreen for different resolutions. (**a**) 0.08-m resolution; (**b**) 0.5-m resolution; (**c**) 1-m resolution; (**d**) 2-m resolution; (**e**) 5-m resolution; (**f**) 10-m resolution.

#### (e) 5-m resolution; (f) 10-m resolution. **4. Discussion**

**Table 5.** Area and percentage of the Fusarium wilt infected regions based on different resolution maps. **Resolution Healthy Area (ha) Diseased Area (ha) Percentage of Diseased Area (%)** CIRE 0.08 m 8.78 6.04 40.8 0.5 m 8.28 6.59 44.3 1 m 8.60 6.28 42.2 2 m 8.38 6.47 43.6 5 m 9.11 5.70 38.5 10 m 9.19 5.69 38.2 CIgreen The results of this study indicate that the CIRE was the optimal red-edge VI and the CIgreen was the optimal non-red-edge VI for developing identification models for banana Fusarium wilt. This is attributed to the fact that as the infection of Fusarium wilt progresses, the chlorophyll content decreases significantly [46], and the CIRE and CIgreen values are sensitive to small variations in the chlorophyll content [37,38]. Furthermore, for the same type of VI, higher OA and Kappa coefficients were obtained for VIs that included the red-edge band than for those without a red-edge band (i.e., CIRE vs. CIgreen and NDRE vs. NDVI). Many studies have demonstrated that the red-edge region is highly sensitive to changes in chlorophyll, and bands in this region are well-suited for estimating the chlorophyll content [47,48], which decreased significantly as the infection of Fusarium wilt progressed. Huang et al. [7] also proved that the red-edge band can be used for disease detection. However, the UAV-based multispectral images used in this study only had five spectral bands, which may not fully reflect the differences in spectral characteristic between healthy and diseased plants. It is necessary to conduct further studies on the sensitivity of certain bands to Fusarium wilt using hyperspectral data.

0.08 m 8.87 5.95 40.1

The results also demonstrated the potential of BLR combined with VIs for the accurate identification of banana Fusarium wilt. This approach provides an ideal framework for using spectral features to determine pathological mechanisms. In this study, the dependent variable was the infection or non-infection of banana Fusarium wilt. BLR is a suitable approach when the predicted variable has a binary nature [32]. Moreover, when the predictor variables are continuous, categorical, or a combination of the two, its performance is better than discriminant analysis [49]. Because BLR is very efficient and highly interpretable and does not require large computational resources, it is a widely used technique to describe the relationship between a dependent variable and multiple independent variables [32]. However, logistic regression is not one of the most powerful algorithms, and some more complex algorithms may easily perform better. Moreover, nonlinear problems cannot be solved with logistic regression due to the linear decision surface. With the development of artificial intelligence, pattern recognition and machine learning methods will become more prevalent for monitoring and forecasting of crop diseases using remote sensing [50].

In this study, VD1 from the Guangxi site and VD2 from the Hainan site were used to verify the Fusarium wilt detection models. The verification results at two locations showed that CIRE and CIgreen had good performances for Fusarium wilt identification with the OA all greater than 70% and Kappa values all greater than 0.4, indicating a good transferability of the Fusarium wilt detection methodology in other areas. However, in Tables 3 and 4, it can be seen that the Kappa values of VD2 were lower than those of VD1. For example, in Table 3, the Kappa value of CIRE was 0.83 in VD1 and 0.59 in VD2, and the Kappa value of CIgreen was 0.74 in VD1 and 0.47 in VD2. This shows that the application of the Fusarium wilt detection methodology in other areas would cause some loss of precision. This situation may be caused by the following factors. First of all, the different varieties of the two experimental sites may be one of the most important reasons affecting the verification results. The variety for VD1 was "Williams B6" and for VD2 was "Baxijiao." There were differences in biophysical and biochemical characteristics between those two varieties. Differences in these variety characteristics can lead to differences in spectral information. Second, due to the differences in planting systems between these two experimental sites, their growth stages differ greatly. When acquiring images in this study, the two experiments were at different growth stages. Besides, it is also better to consider factors, such as planting density, soil types, and crop growth environmental conditions, that could affect the applicability of the Fusarium wilt detection methodology. Therefore, when applying the method in different regions, it is suggested to optimize the parameters of BLR if there is a large difference between the application and the modeling area of banana planting and growth.

This study demonstrates that UAV-based multispectral imagery is well-suited for the identification of banana Fusarium wilt disease. We also simulated the resolutions of satellite-based imagery (i.e., WorldView series with a resolution of 0.5 m, GF-2 with a resolution of 1 m, GF-1 and GF-6 with a resolution of 2 m, RapidEye with a resolution of 5 m, and Sentinel-2 with a resolution of 10 m) to assess the effects of imagery with different spatial resolution on the identification of disease. The results showed that imagery with a spatial resolution higher than 2 m had good identification accuracy of Fusarium wilt, which might be related to the plant spacing of bananas. As the resolution decreased, the mixed pixel effect worsened, and the monitoring accuracy decreased. However, the resolution was not the only difference among the UAV and satellites. The satellites captured information in wavelengths that was not the same as the ones used in the UAV sensors. Hence, the simulated results of the different resolutions need to be further verified while applied with the actual satellite data. In this study, single date multispectral imagery was used, which represents limitations with regard to determining the spectral response mechanism of the changes in the biophysical and chemical parameters caused by Fusarium wilt. In the future, multitemporal and hyperspectral imagery should be investigated. Moreover, the differences in the spectral response characteristic between Fusarium wilt and other yellowing phenomena caused by other stresses (i.e., nutrition deficiency and drought stress) should also be examined.
