Classification of white root disease severity levels using rubber tree spectral data

In Southeast Asia, white root disease (WRD) is one of the most damaging diseases for rubber plantations, which can cause up to 50% reduction in natural rubber productivity. The disease can be controlled if it is detected earlier, where affected trees are treated immediately. Early detection means detecting the disease before any visible symptoms of infection appear (pre-symptomatic phase). However, the existing detection technique is labor-intensive, costly, and time-consuming because it depends on skilled personnel and laboratory analysis. Therefore, there is a need for an accurate and reliable in-situ measurement system to detect the disease at an early stage. Thus, the objective of this study was to explore the capability of a low-cost visible shortwave near-infrared (VSNIR) in combination with machine learning as an early detection method for WRD from rubber trees. A total of 134 rubber trees at different severity levels namely, healthy (S0), light (S1), moderate (S2), severe (S3), and very severe infection (S4) rubber trees were used in the study. Samples of leaves, latex, and soil around rubber trees were collected from both healthy and affected trees. For the leaf samples, a visible shortwave near-infrared (VSNIR) spectrometer was used to record the spectral data. At the same time, SPAD value was measured using a SPAD meter, moisture content (MC) was measured using a moisture content analyzer, and leaf color was measured using a colorimeter. For the latex samples, the total soluble solid content (TSC) was determined using standard methods ISO 124:2001 in the laboratory. Finally, for the soil samples, the pH and moisture content (MCS) properties around the rubber trees were determined using a Portable Soil Analyzer with multiprobes. The calibration and prediction models were developed to correlate the spectral data with the chemical properties of the samples using the partial least square (PLS) regression method. Feature selection algorithms such as ANOVA, Chi-Square and RELIEF were used to select the most significant wavelength for classification models. Then, four classifiers, namely, artificial neural networks (ANN), support vector machine (SVM), random forest (RF), and k-nearest neighbor (kNN) were applied to classify spectral data into different severity levels of WRD. The result shows that the leaf sample from S4 had the lowest SPAD value of 47 and the lowest MC of 41.95 % w.b. The result presented SPAD value and MC, which indicates an inverse relationship with the severity levels of WRD. The L* value increased from S0 (32.97 ± 3.37) to S4 (38.43 ± 4.97), while the a* value was decreased from S0 (-1.58 ± 1.92) to S4 (-2.16 ± 3.04) and the b* value was increased from S0 (12.64 ± 3.44) to S4 (17.79 ± 6.39). The result shows that the leaves are lighter, greener, and have a more yellowish appearance with increasing disease severity. For the soil samples, the average pH value remains relatively consistent across all severity levels, ranging from 6.20 (S0) and 6.40 (S4), while MCS decreased as the disease severity increased. The MC value dropped significantly from S0 (33.50 % w.b.) to S4 (21.70 % w.b.). For the latex samples, the TSC value was decreased from S0 (1.90%) to S3 (1.58%), this indicates that as the severity of WRD increases, the latex becomes more diluted. The crop parameters obtained from the samples at different severity levels significantly difference at (p<0.05) except for soil samples. In term of PLS regression models, the results showed that the prediction of chlorophyll content from the leaf sample had the highest coefficient determination (R 2 ), with a calibration model value of 0.99 and a prediction model value of 0.99. For the MC prediction of rubber leaf samples, the calibration model gave an R 2 value of 0.61, whereas the prediction model gave an R 2 value of 0.65. For the color prediction of the rubber leaf samples, R 2 values for calibration models for L*, a*, and b* were 0.88, 0.85, and 0.89, respectively. For the prediction models, R 2 values for L*, a*, and b* were 0.95, 0.95, and 0.94, respectively. The result showed that the SPAD value of the leaf samples had higher R2 than other crop parameters. Thus, SPAD value could be the parameter that VSNIR can predict across all severity levels. The RELIEF feature selection method was the most effective for identifying the significant wavelengths that distinguish between different disease severity levels. Additionally, the kNN classifier was the most accurate for classifying WRD severity, enabling reliable early disease detection. In conclusion, utilising VSNIR is a technology that shows promising in detecting WRD in rubber trees at an early stage by analysing changes the SPAD value, leaf color and MC levels. Therefore, VSNIR technology offers possible applications for controlling the disease's spread within the farm.

Text IKP 2025 1 - Full Text.pdf Available under License Creative Commons Attribution Non-commercial No Derivatives. Download (4MB) Text IKP 2025 1.pdf Restricted to Repository staff only Available under License Creative Commons Attribution Non-commercial No Derivatives. Download (7MB)

Actions (login required)

View Item