Large-Scale Monitoring of Potatoes Late Blight Using Multi-Source Time-Series Data and Google Earth Engine

Abstract

Effective monitoring and management of potato late blight (PLB) is essential for sustainable agriculture. This study describes a methodology to improve PLB identification on a large scale. The method combines unsupervised and supervised machine learning algorithms. To improve the monitoring accuracy of the PLB regression model, the study used the K-Means algorithm in conjunction with morphological operations to identify potato growth areas. Input data consisted of monthly NDVI from Sentinel-2 and VH bands from Sentinel-1 (covering the year 2021). The identification results were validated on 221 field survey samples with an F1 score of 0.95. To monitor disease severity, we compared seven machine learning models: CART decision trees (CART), Gradient Tree Boosting (GTB), Random Forest (RF), single optical data Random Forest Time series model (TS–RF), single radar data Random Forest Time series model (STS–RF), multi-source data Gradient Tree Boosting Time series model (MSTS–GTB), and multi-source data Random Forest Time series model (MSTS–RF). The MSTS–RF model was the best performer, with a validation RMSE of 20.50 and an R² of 0.71. The input data for the MSTS–RF model consisted of spectral indices (NDVI, NDWI, NDBI, etc.), radar features (VH-band and VV-band), texture features, and Sentinel-2 bands synthesized as a monthly time series from May to September 2021. The feature importance analysis highlights key features for disease identification: the NIR band (B8) for Sentinel-2, DVI, SAVI, and the VH band for Sentinel-1. Notably, the blue band data (458–523 nm) were critical during the month of May. These features are related to vegetation health and soil moisture are critical for early detection. This study presents for the first time a large-scale map of PLB distribution in China with an accuracy of 10 m and an RMSE of 26.52. The map provides valuable decision support for agricultural disease management, demonstrating the effectiveness and practical potential of the proposed method for large-scale monitoring.

1. Introduction

Potatoes are recognized for their reduced greenhouse gas emissions and water requirements, conferring substantial environmental advantages over other key agricultural crops [1]. In light of these benefits, China has implemented initiatives to establish potatoes as a dietary staple, with the goal of reducing the disparity in production yields in comparison to leading producers. It is anticipated that these measures could lead to a reduction in the carbon-land-water impact of staple crops by 17–25% by 2030 [1,2]. Late blight, induced by the pathogen Phytophthora infestans, stands as a principal disease affecting potato cultivation, resulting in considerable losses in yield [3,4], and thus, impacting the role of potatoes in sustainable development endeavors. Consequently, extensive mapping of the prevalence of potato late blight (PLB) is essential for crisis management, facilitating effective surveillance, and achieving optimal disease management strategies. However, traditional disease monitoring is primarily dependent on field surveys, which are laborious and time-consuming, and are not aligned with the real-time monitoring requirements for extensive regions [5].

Given the limitations of traditional monitoring methods, advanced technologies are being increasingly explored. Previous studies have shown that, at the leaf organ level and under laboratory conditions, hyperspectral imaging technology can be used to achieve early detection of PLB, even in the invisible stage of leaf disease, and hyperspectral monitoring can be achieved non-destructively [6,7,8]. After remote sensing spectroscopy proved to be useful for the study of PLB, subsequent studies mostly combined this with UAV for monitoring. For example, some studies have used drones for large-scale surveillance of PLB. One approach to assessing PLB severity involves using RGB images captured by drones [9]. Another method evaluates the extent of late blight using high-resolution multispectral UAV images combined with machine learning (ML) algorithms [10]. A deep learning model known as CropdocNet has been proposed for accurately and automatically diagnosing late blight from UAV hyperspectral images [11]. Additionally, research by Sun et al. (2023) explores collaborative PLB monitoring using multiple drone sensors [12]. These studies demonstrate the feasibility of using remote sensing spectral information combined with ML or deep learning algorithms to detect PLB [12]. These methods include the use of multi-layer perceptrons, convolutional neural networks (CNNs), support vector regression (SVR), and Random Forest (RF) applied to multi-spectral UAV imagery. The results indicate that ML algorithms can achieve high accuracy, suggesting their potential to replace traditional visual estimation methods with an acceptable mean absolute error of 11.72% [12]. Despite the success of UAV technology and ML algorithms, their ability for large-scale disease monitoring remains limited in continuity and coverage. In contrast, satellite remote sensing offers an economical and efficient method for large-scale characterization of ground objects and long-term change detection [13], making it a valuable tool for disease monitoring. The existing studies of PLB monitoring by satellite remote sensing have confirmed that the relationship between spectrum and PLB has scale variability, but is feasible and has great potential for the use of Sentinel-2 multispectral monitoring of PLB [14].

Furthermore, crop distribution data serve as the foundational information for precise disease monitoring. Currently, there are few methods and models for mapping potato distribution; nevertheless, inspiration can be drawn from summarizing identification algorithms for other crops [15,16,17,18,19]. The Google Earth Engine (GEE) platform, in conjunction with Sentinel-1/2 data, effectively utilized phenology data and ML techniques to create continuous, inter-annual distribution maps of oilseed rape with a 10-m resolution [20]. Object-based methods and deep learning algorithms were employed to provide a novel perspective on the mapping of crop types in the northwest Ardabil region of Iran [21]. A multi-temporal Gaussian mixture model developed on the GEE platform was applied to efficiently map maize distributions [22]. Higher resolution of corn and soybean mapping in China [23] was achieved by PlanetScope data and a RF-based method. These successful cases mainly utilize data from optical satellites, which offer high spatial and temporal resolution, covering the visible, near-infrared (NIR), and shortwave-infrared (SWIR) bands. These bands can capture different spectral characteristics of crops, which are helpful for crop classification and monitoring [24,25].

While optical remote sensing images have garnered considerable attention, the full potential of Synthetic Aperture Radar (SAR) technology remains largely untapped. SAR sensors can obtain images under various weather conditions, making them ideal tools for long-term and multi-seasonal monitoring. In addition, SAR holds extensive potential for various applications in agricultural monitoring [26], for it can provide important information about crop growth and soil moisture conditions [25]. Variations in radar backscatter during different growth seasons can also aid in detecting crop types, as different crop growth stages produce distinct scattering mechanisms, including surface scattering, corner scattering, and volume scattering. Notably, the RADARSAT satellite derived C-band SAR data have exhibited superior capability in differentiating rice and potato crops within the Ganges Plain. This methodology has attained remarkable classification accuracy, reaching 94% for rice and 93% for potatoes during the initial phases of their developmental cycle [27]. By fusing multi-temporal Sentinel-1 SAR data and Sentinel-2 optical data, and applying the 3D U-Net deep learning network, scientists have significantly improved the accuracy of crop type mapping, providing an efficient crop monitoring and classification method for data-driven agriculture [28]. SAR and multispectral data have been successfully used to extract crop-specific land phenology from major crops in Europe. Validation against ground truth data confirmed the accuracy, demonstrating the complementary nature of these satellite data types in delivering crop phenology insights, which bolsters the precision and dependability of crop surveillance [29]. These successful cases of crop mapping using time series multi-source data and phenological information provide solid foundations for PLB distribution mapping [30,31].

Moreover, conventional techniques for processing remote sensing data for large-scale analyses are intricate, involving complexities across spatial and temporal dimensions [32]. The task of large-scale crop mapping entails the processing and handling of extensive volumes of diverse satellite imagery sourced from multiple sensors, which poses considerable challenges associated with ’big data’. As remote sensing technology advances, monitoring crop disease using large volumes of satellite imagery has become feasible, especially with the emergence of the GEE platform. GEE provides powerful capabilities for big data analysis and access to high-resolution imagery on a global scale. It has shown excellent performance in accessing remote sensing products and pre-processing through the cloud platform [33,34], enabling extensive surveillance and cartographic representation of PLB [35]. The unrestricted access to satellite-based global observations offers a supplementary method to consistently acquire large-scale PLB information at a low cost, both spatially and temporally, providing a new solution for potato production management [36,37]. Despite previous success in crop monitoring and distribution mapping based on GEE satellite remote sensing technology, there is limited research on the identification and mapping of large-scale crop disease. Most existing studies rely on UAV-scale hyperspectral and multispectral modeling for PLB monitoring, thereby restricting the utilization of reliable and freely available satellite data for large-scale and efficient disease monitoring.

At present, there is a growing trend in the utilization of remote sensing technology for agricultural monitoring, but challenges remain in large-scale detection and spatial visualization of PLB [14,18]. Moreover, there is no remote sensing-based potato distribution mapping in China, and only limited research on large-scale PLB monitoring methods. This study aims to develop a large-scale PLB monitoring method based on GEE to improve the accuracy and monitoring efficiency of disease identification, so as to provide real-time information support for agricultural disease management. The specific objectives are (1) precise identification of potato planting areas by using multi-source time series remote sensing data and K-Means clustering algorithm; (2) evaluation of the application effect of Sentinel-1 and Sentinel-2 data fusion in disease surveillance to reveal the complementarity of different data sources; and (3) provision of PLB monitoring framework based on machine learning algorithm, in order to provide a scientific basis for sustainable agricultural development and to reduce the yield loss caused by PLB.

2. Materials and Methods

2.1. Study Area

The research was undertaken in Guyuan County of Hebei Province and Duolun County of the Inner Mongolia Autonomous Region, both of which are significant potato-producing regions within China. The two regions are characterized by a diverse agricultural landscape, with main crops including potatoes, wheat, and corn. As depicted in Figure 1, the research area is situated within a temperate continental grassland climate region, with high temperatures in summer, abundant rainfall, and high humidity. In the study area, potato cultivation and sprouting occur between April and May, followed by a period of rapid growth from June to August, with the harvesting season extending from late August to September. The humid climatic conditions create an environment conducive to the development of PLB [9,32]. Throughout the history of potato cultivation, PLB has consistently been a major disease affecting yield and quality. Caused by the fungus Phytophthora infestans [38,39,40], PLB has historically caused severe yield losses in these areas, exerting a significant negative impact on the local agricultural economy.

2.2. Data Collection and PLB Severity Assessment

From 11–13 August 2021, a field survey of PLB was conducted in the study area to obtain first-hand disease occurrence data and lay the foundation for subsequent analysis and modeling.

In the study area, 221 representative samples were carefully selected to ensure that the samples could comprehensively cover different planting areas and disease occurrence. Using high-precision GPS equipment, the geographical position coordinates of each sample point are accurately recorded, which provides a reliable basis for the spatial analysis of subsequent data and matching with remote sensing images.

The observation area of each sample point is set to be consistent with the pixel size of the Sentinel-2 satellite remote sensing image, that is, 10 m ×10 m. This setting helps to achieve accurate correspondence between ground survey data and remote sensing data, improve the accuracy of data fusion analysis, and ensure that disease monitoring results based on remote sensing images better reflect the actual situation.

In the course of field investigation, the PLB visual estimation method was used to assess the degree of disease occurrence in each sample site in detail. According to the characteristics of the disease on the plant, the severity of PLB was divided into several grades, as shown in

Table 1. Methods for identifying the degree of PLB in field measurement.

PLB Severity (%)	Description
0.0%: No symptoms	No disease symptoms observed.
0.1%: Initial infection	Initial sporangium lesions were detected.
1.0%: Mild infection	Each plant has 5–10 plaques.
5.0%: Moderate infection	1/10 of the leaves were affected, with 50 spots per plant.
25.0%: High infection rate	Almost every leaf fragment is affected and the plant morphology is normal.
50.0%: Severe infection	Each plant is affected and 50% of its leaf area is destroyed.
75.0%: Advanced infection	75% of the leaf area is destroyed and the field color is changed.
95.0%: Severe wilting	There are only a few leaves left on the plant.
100.0%: Complete wilting	The leaves are completely necrotic and the stems are dead or dying.

[9,39,41]. By on-site observation and the experienced judgment of professionals, the PLB severity level of each sample point was determined, so as to obtain accurate ground disease occurrence data, and provide key reference information for subsequent model training and verification. Table 1 lists nine discrete disease severity levels (e.g., 0.0%, 0.1%, 1.0%, 5.0%, 25.0%, etc.) that were used in the field to assess and categorize disease severity during the data collection process. While the continuous output of the PLB prediction model can be mapped to these discrete levels for interpretation and validation, the actual output is not limited to these specific values. Instead, it provides a finer granularity of disease severity.

2.3. Multi-Resource Remote Sensing Data

In this research, Sentinel-1 and Sentinel-2 data were merged on GEE to generate a 2021 monthly time series for mapping potato distribution and monitoring PLB. The data spanned January to December 2021.

Selected Sentinel-2 Level-2A data (ID: “COPERNICUS/S2_SR_HARMONIZED”) underwent calibration and correction, totaling 436 scenes. All bands were utilized, with QA60 for cloud and shadow masking. Monthly cloud-free composites were created. Sentinel-2’s high resolution and red edge bands are vital for crop classification and disease monitoring. Sentinel-1 SAR data (ID: “COPERNICUS/S1_GRD”) were processed for backscatter features in VV and VH polarizations. The COPERNICUS-provided dataset included C-band images at 5.405 GHz. Interferometric wide mode images were selected, and spatial filtering ensured ROI relevance. Noise pixels below −30 dB were masked. Composites were created from 107 scenes within the study period.

2.4. Research Methods

This research develops a scalable PLB detection and mapping approach using GEE (Figure 2), enhancing disease surveillance precision. The study created a year-long dataset by integrating Sentinel-1 and Sentinel-2 data, extracting various indices to analyze spatial patterns and complexity by time series dataset. The detection of PLB involves multi-source remote sensing data, which have high-dimensional and complex feature distribution. K-Means clustering can effectively preprocess data and cluster similar samples together, thereby reducing the complexity of the data and identifying the spatial distribution of potatoes, so as to improve the efficiency of subsequent classification. Firstly, K-Means clustering was used to identify the main patterns and structures in the data, so as to provide more representative training samples for Random Forest. As a powerful ensemble learning method, Random Forest is able to process high-dimensional data and automatically select important features, thereby improving the accuracy and robustness of classification. A combination of K-means clustering and RF algorithms was applied for potato identification and PLB detection, with a decision tree model built by field data to boost model accuracy. The model was evaluated using F1 score, the coefficient of determination (R²) and the root mean square error (RMSE), yielding spatial distribution of PLB and detailed severity maps.

2.4.1. Construction of Features and Indices

In this study, optical remote sensing and radar remote sensing data are used. The indices calculated based on these original bands are as follows.

Selecting appropriate vegetation indices is essential for quantifying plant responses to disease-induced stress, which manifests in various physiological and biochemical changes [15,42]. In this study, a variety of vegetation indices, moisture indices, soil and building indices and other relevant indices were comprehensively used to capture and analyze potato growth status and late blight (PLB) occurrence. This includes Normalized Vegetation Index (NDVI) [43,44] and Ratio Vegetation index (RVI) [45], which are used to evaluate the health and growth of vegetation. Water indices, such as the Normalized Humidity Index (NDWI) and Disease Water Stress Index (DWSI) [44,46,47] to detect water conditions and identify disease-related water stress; soil and building indices, such as the Soil Adjusted Vegetation Index (SAVI) [48] and the Normalized Building Index (NDBI), which are used to consider soil impacts and differentiate urban areas; and other indices, such as Normalized Canopy Chlorophyll Index (NDCI) [49], were used to estimate chlorophyll content and further analyze the physiological characteristics of vegetation. The use of these indices aims to make full use of multi-band information for remote sensing data, accurately identify and monitor the occurrence and development of PLB and provide a scientific basis for disease management and agricultural decision-making. The indices utilized are enumerated in

Table 2. List of indices calculated based on Sentinel-2 optical bands.

Remote Sensing Index	Description
$NDVI={\displaystyle \frac{NIR-RED}{NIR+RED}}$	Normalized vegetation index, reflecting vegetation growth [43,44]
$NDWI={\displaystyle \frac{GREEN-NIR}{GREEN+NIR}}$	Normalized moisture index for water detection [50]
$NDBI={\displaystyle \frac{SWIR-NIR}{SWIR+NIR}}$	Normalized building index for urban area detection [51]
$SAVI={\displaystyle \frac{(NIR-RED)\times (1+0.5)}{NIR+RED+0.5}}$	Soil adjustment vegetation index, taking into account soil influences [48]
$IBIA=2\times {\displaystyle \frac{SWIR1}{SWIR1+NIR}}$ $IBIB={\displaystyle \frac{NIR}{(NIR+RED)}}+{\displaystyle \frac{GREEN}{GREEN+SWIR1}}$ $IBI={\displaystyle \frac{IBIA-IBIB}{IBIA+IBIB}}$	Construction index, for detection of urban areas [52]
$RVI={\displaystyle \frac{NIR}{RED}}$	The ratio vegetation index reflects the relative health status of vegetation [45]
$DVI=NIR-RED$	The Difference Vegetation Index, which represents simple differences in vegetation [45]
$NDCI={\displaystyle \frac{NIR-RED}{NIR+RED}}$	Normalized canopy chlorophyll index to estimate chlorophyll content [49]
$DWSI2={\displaystyle \frac{B6-GREEN}{SWIR1+RED}}$	Disease water stress index is very sensitive to changes in leaf pigment, internal leaf structure and moisture content [44,46,47]
$DWSI3={\displaystyle \frac{B7-GREEN}{SWIR1+RED}}$	Disease water stress index, which is the similar as DWSI2
$DWSI4={\displaystyle \frac{B8A-GREEN}{SWIR1+RED}}$	Disease water stress index, which is the similar as DWSI2

.

Furthermore, radar features, derived from Sentinel-1 SAR imagery, aid in assessing vegetation biomass, moisture content, and surface roughness, which are critical for monitoring crop health and detecting diseases. The study leverages Sentinel-1 SAR imagery’s distinct interaction with Earth’s surface to glean unique insights into surface and vegetation characteristics [28,53]. It specifically employs two polarization modes: Vertical Transmit and Vertical Receive (VV) and Vertical Transmit with Horizontal Receive (VH). The backscattering coefficients, ${\gamma }_{VV}$ and ${\gamma }_{VH}$, are calculated from the SAR data to evaluate the radar response [35,54], defined as the ratio of the backscattered power to the incident power:

$${\gamma }_{VV}={\displaystyle \frac{P_{backscattered,V}}{P_{incident}}}$$

(1)

$${\gamma }_{VH}={\displaystyle \frac{P_{backscattered,VH}}{P_{incident}}}$$

(2)

Here, the polarization modes are denoted as $P_{vv}$ and $P_{VH}$, respectively.

In addition, the texture features can quantify the spatial complexity by assessing the statistical relationships between pixel pairs in a given direction [55]. The Gray Level Co-occurrence Matrix (GLCM) method is an effective technique for texture analysis [56]. The GLCM was employed to extract texture indices, such as contrast, variance, mean, and difference variance, which enhance the analytical capabilities of images and provide essential supplementary information for land cover classification and vegetation assessment. Utilizing the robust computational power of the GEE platform with a 3 × 3 window size and a 0° orientation, these indices were calculated for obtaining a set of features that characterize the texture of the study area.

2.4.2. A Potato Identification Model Based on Multi-Source Time-Series Remote Sensing Features and Clustering Algorithms

To improve the PLB identification accuracy, the study utilized the K-Means clustering algorithm to identify potato growing areas because of its ability to efficiently process large-scale data by integrating multi-source time-series remote sensing features in an unsupervised clustering task. The K-Means clustering algorithm is a distance-based clustering method [57,58,59,60]. The algorithm functions through a recursive reduction of the Euclidean distance between each datum and its respective cluster centroid. Its widespread application is attributed to its simplicity, intuitive design, and efficacy. The objective is to segregate data points into clusters in a manner that optimizes for minimal intra-cluster variance, which translates to minimizing the aggregate Euclidean distances from each data point to the centroid of its designated cluster. The objective function can be expressed as:

$$J={\displaystyle \sum _{k=1}^K{\displaystyle \sum _{x\in S_k}{‖x-{\mu }_k‖}^2}}$$

(3)

where $J$ is the objective function, $K$ is the number of clusters, which is the set of points in the $k$-th cluster, and $x$ is a data point within the cluster, which is the center of the $k$-th cluster, representing the Euclidean distance.

In order to efficiently utilize this clustering method, we created a multi-temporal dataset spanning from January to December 2021 using Sentinel-2 optical and Sentinel-1 radar images. The GEE platform facilitated pre-processing, including noise reduction and smoothing. NDVI and VH band data were extracted to quantify vegetation health and crop structure. These indices were merged into a multi-band image for K-Means clustering, which analyzed 10,000 randomly selected pixels to identify potato-growing areas. The results were validated with 221 field samples labeled as potato or non-potato areas. This integrated approach provides a robust framework for precision agriculture.

In this study, K-means unsupervised classification was first employed to identify potato planting locations, thereby enhancing the accuracy of subsequent machine learning regression. To avoid errors from the initial unsupervised method affecting the accuracy of the subsequent machine learning regression, we implemented several measures to ensure the reliability of our results.

We used the K-means clustering method to identify potato-growing areas by integrating multi-source time-series remote sensing features with field sample data. Morphological operations were then performed on the clustering results to optimize isolated, broken pixels (which are unlikely to represent entire fields). Using the GEE platform, we processed and analyzed field grid images through a series of image processing and morphological analysis methods to identify and optimize crop distribution. Specifically, we defined a 5 × 5 structural element (K) and applied morphological opening operations, including erosion and dilation, to each pixel to remove noise from the image. Gaussian filter is applied to smooth the image after open operation to further reduce the noise. Then, a morphological closure operation is performed to connect adjacent areas by first expanding and then corroding. After that, the number of pixels in each connected region of the image after the closed operation is calculated to extract the features. Finally, the area threshold T is set to keep only the area greater than or equal to T, thereby streamlining and optimizing the classification of farmland areas. This method effectively improves the accuracy and efficiency of farmland distribution identification. In addition, we carried out a detailed error analysis on the clustering results of the K-Means method and selected four real scenes from Google Earth images to illustrate the spatial details of classification, so as to identify possible areas of classification error. Finally, a result-oriented test is carried out. In the final selected model, the correctness of the model is evaluated on the independent verification set, assuming that there are classification errors that may be generated by the K-Means method.

2.4.3. Euclidean Distance-Based Determination of Separability of PLB Characteristics

Previous studies have shown that the spectral curve of Sentinell-2 can still distinguish and identify PLB to varying degrees at 10 m spatial resolution [14]. This investigation employed Sentinel-2 imagery, initially processed via the GEE platform, to evaluate spectral variations across various Pineapple Leaf Blight (PLB) severity classes, thereby enhancing the comprehension of machine learning (ML) algorithms in PLB surveillance. The initial phase entailed the acquisition of data spanning from 6–15 August 2021, which was subsequently subjected to precise temporal and spatial filtering to ascertain the pertinence and precision of the dataset. Thereafter, high-resolution imagery was consolidated through image fusion techniques to ensure the maintenance of spatial and temporal coherence within the area of interest. To obtain surface feature samples, 221 sample regions from the synthesized images were extracted based on predefined feature coverage labels that included information on the severity of PLB. The samples were subsequently converted into feature vectors and mean and covariance matrices were computed for each PLB class. Additionally, the Euclidean Distance between the sample mean vectors was calculated to quantify the spectral differences between different PLB classes. This distance measurement offers an evaluation criterion to assess the separability between different classes. A smaller Euclidean Distance suggests similarity in spectral features, whereas larger distances denote better separability.

2.4.4. A PLB Identification and Spatial Visualization Model Based on ML and Time-Series Remote Sensing Features

In recent years, machine learning (ML) methods have been extensively applied to address the issue of PLB [11,14,17,61,62]. At the same time, unlike complex deep learning models that require a large number of training samples and parameters, GEE has some convenient built-in ML algorithms that allow users to access and process petabytes of remote sensing data. Using these advantages to extract phenological information from time series remote sensing data, favorable insights into the development of different diseases can be obtained from phenological data extracted from each crop growth cycle [34,63,64].

The course of PLB has obvious seasonal and temporal dynamics [65]. By using time series data, changes in the disease at different stages of growth can be captured [41,66,67], allowing for more accurate prediction and monitoring of PLB severity. Sentinel-1 SAR data and Sentinel-2 optical data each carry unique information. SAR data can obtain surface information under cloudy and rainy conditions, while optical data provides rich spectral information [68]. By fusing these two types of data, multi-dimensional characteristics of potato growth and disease development can be more fully captured. In particular, PLB changes manifest differently in the early and late stages of the disease, and these changes can be better identified through time series analysis of multi-source data [43,69,70]. Random Forest (RF) [71] is a powerful ensemble learning method capable of processing high-dimensional data and automatically selecting important features. It performs well in the processing of time series data and can effectively reduce overfitting and improve the generalization ability of the model. Therefore, we propose MSTS–RF (Time Series Random Forest with Multi-Source Data) method. In order to verify the reliability and superiority of the model, we will conduct several controlled experiments, specifically. The first is to verify the superiority of different ML models in PLB problems, and find the most effective model to solve PLB monitoring; the second is to compare the performance of the model without using time series data and the model with time series data; the third is to compare the performance differences between the two single time series data models of radar and optical data and the multi-source time series models, respectively. Finally, the optimal model is selected as the monitoring model for this study. The specific model introduction and modeling process are as follows:

Classification and Regression Trees (CARTs): CARTs employ decision trees for both classification and regression tasks [72]. In this study, we use the ‘ee.Classifier.smileCart()’ method on the GEE platform to create a CART classifier. In order to optimize the hyperparameters, we search in the range of ‘maxNodes’ (maximum number of leaf nodes) and ‘minLeafPopulation’ (minimum number of samples per leaf node) through grid search and finally determine the number of ‘maxNodes’ to be 10. ‘minLeafPopulation’ is 1 as the best parameter. Training data sets account for 70% and validation data sets account for 30%. The input features include various vegetation indices (such as NDVI, NDWI, NDBI, SAVI, IBI, RVI, DVI, NDCI, DWSI2, DWSI3, DWSI4) and texture features (such as contrast, variance, and standard deviation) calculated based on Sentinel-2 images synthesized one week before and after the date of 2021-08-12, as well as the VV and VH bands from Sentinel-1. The label attribute ‘PLB_percen’ indicates the severity of PLB. During model training, we use root mean square error (RMSE) and coefficient of determination (R²) to evaluate model performance.

Gradient Tree Boosting (GTB): This ensemble technique progressively incorporates trees to rectify the residuals of the preceding model, directed by the gradient of the loss function [73]. In this study, we use the GEE platforms ‘ee.Classifier.smileGradientTreeBoost()’ method to create a GTB regression model. The training process is similar to CART. Training and validation are separated, with 70% of the data used for model training and 30% for validation.

Random Forest (RF): An ensemble technique that constructs multiple decision trees and amalgamates their predictive outcomes [71], introducing randomness to avoid overfitting and enhance accuracy. In this study, we use the GEE platform’s ‘ee.Classifier.smileRandomForest()’ method to create a GTB regression model. The training process is similar to CART. Training and validation are separated, with 70% of the data used for model training and 30% for validation.

Time Series Random Forest Regression with Single Optical time series data (TS–RF): A specialized RF model that uses time-series data from a single source, Sentinel Optics, to analyze and predict trends based on temporal patterns. Using Sentinel-2 satellite data, vegetation and surface indices, such as NDVI, NDWI, NDBI, SAVI, IBI, RVI, DVI, NDCI and DWSI, were calculated, and rich surface features were extracted. In addition, the spatial information is enhanced by GLCM texture analysis. A multi-band time series dataset was established by synthesizing images month by month from May to September 2021. The image features were then trained using an RF model for monitoring PLBS. In the training phase, the data set is randomly divided into a training set accounting for 70% of the data and a validation set accounting for 30% of the data to evaluate the performance of the model. In order to optimize the hyperparameters, we search in the range of ‘numberOfTrees’ (the number of decision trees to create) through grid search, and finally determine ‘numberOfTrees’ 10 as the best parameter. The label attribute ‘PLB_percen’ indicates the severity of PLB. During model training, we use root mean square error (RMSE) and coefficient of determination (R²) to evaluate model performance.

Time Series Random Forest Regression with radar time series data (STS–RF): A specialized RF model that uses time series data from a single source from radar to analyze and predict PLB trends based on time patterns. It is similar to the TS–RF model in construction, but only the VV and VH bands synthesized every month are used for training to explore the disease monitoring capability of the radar.

Time Series Random Forest Regression with Multi-source Data (MSTS–RF): This study presents a Time Series Random Forest Regression model with Multi-source Data (MSTS–RF) to monitor PLB severity using remote sensing data. The dataset consists of 200 bands synthesized from Sentinel-2 and Sentinel-1 satellite data, covering all Sentinel-2 bands, remote sensing indices (e.g., NDVI, NDWI, NDBI, etc., see Table 2), texture features from GLCM, and radar polarization bands (VV and VH). Monthly image composites were generated for May-September on a month-by-month basis to create a multi-band time series dataset. The training data, labeled with PLB severity classes (0–100%) based on field surveys (221 samples), were partitioned into a 70% training set and a 30% validation set. The Random Forest model was configured with the following hyperparameters: 1000 trees, default variables per split (square root of total variables), minimum leaf population of 1, and a bag fraction of 0.5. The maximum number of nodes per tree was not limited, and the random seed was set to 0 to ensure repeatability. Model performance was evaluated using R² and RMSE. The MSTS–RF outputs continuous values representing PLB severity and spatial distribution, visualized as image layers with legends for reference.

Time Series Gradient Tree Boosting Regression with Multi-source Data (MSTS–GTB): In order to verify the performance of multi-source time series data in different machine learning models, it is proposed to refer to the MSTS–RF model training method, which is consistent except for the different machine learning methods based on it. Using ‘ee. Classifier.smileGradientTreeBoost (10)’ training, we utilized R² to assess model fit and RMSE to measure the deviation between model predictions and actual data.

Finally, the optimal model is selected comprehensively through the verification performance of these seven models.

2.4.5. Performance Metrics

To evaluate the effectiveness of the proposed models and methods for detecting and monitoring PLB, several performance metrics were employed. These metrics provide quantitative measures of model accuracy, precision, and overall predictive power, allowing for a comprehensive assessment of the model’s performance. The key performance metrics used in this study include $Accuracy$, $Precision$, $Recall$, $F1$, Root Mean Square Error (RMSE), and Coefficient of Determination (R²). Their calculations are as follows.

$Accuracy$: refers to the ratio of the number of samples correctly predicted by the model to the total number of samples.

$$Accuracy={\displaystyle \frac{TP+TN}{TP+TN+FP+FN}}$$

(4)

$Precision$: refers to the proportion of plants that the model predicts to be affected by late blight that are actually affected.

$$Precision={\displaystyle \frac{TP}{TP+FP}}$$

(5)

$Recall$: refers to the proportion of plants that the model correctly detects for late blight. It tells us how successfully the model can predict how many plants will actually be infested.

$$Recall={\displaystyle \frac{TP}{TP+FN}}$$

(6)

$F1$ : The F1 score is a blended average of precision and recall, which combines the accuracy and recall of the model.

$$F1={\displaystyle \frac{Precision\times Recall}{Precision+Recall}}$$

(7)

Here, $TP$ (True Positive) is the number of true examples, $TN$ (True Negative) is the number of true negative examples, $FP$ (False Positive) is the number of false positive examples, and $FN$ (False Negative) is the number of false negative examples.

The R-square (coefficient of determination), Root Mean Square Error (RMSE) and Mean Square Error (MSE) were chosen to evaluate the goodness-of-fit of the regression model as follows:

$$R^2=1-{\displaystyle \frac{{\displaystyle \sum _{i=1}^n{(y_i-{\widehat{y}}_i)}^2}}{{\displaystyle \sum _{i=1}^n{(y_i-\overline{y})}^2}}}$$

(8)

$$RMSE=\sqrt{{\displaystyle \frac{1}{n}}{\displaystyle \sum _{i=1}^n{(y_i-{\widehat{y}}_i)}^2}}$$

(9)

where ${\mathrm{y}}_{\mathrm{i}}$ is the observed value, ${\widehat{\mathrm{y}}}_{\mathrm{i}}$ is the model prediction, and n is the sample size.

To intuitively evaluate the performance of the selected optimal model under different training set sizes and prevent overfitting, we generated learning curves by randomly selecting samples of varying proportions from the original data to form training and validation sets. We tested several ratios, including 90% training and 10% validation, 70% training and 30% validation, 50% training and 50% validation, 30% training and 70% validation, and 10% training and 90% validation. For each ratio, the raw data were divided into training and validation sets, and the selected optimal model was trained on each training set size. We then calculated the performance metrics—root mean square error (RMSE) and coefficient of determination (R²)—for both the training and validation sets. RMSE measures the difference between the model’s predicted values and the actual values, while R² indicates how well the model fits the data. Using Python 3, we plotted learning curves with the number of training samples on the horizontal axis and the performance metrics (RMSE and R²) on the vertical axis, drawing separate curves for the training and validation sets.

3. Results

3.1. Spatial Clustering Analysis Based on K-Means Algorithm and Sentinel 1/2 Time-Series Data

The monthly time series data from Sentinel-1 and Sentinel-2 satellites were processed within the study area throughout 2021. The K-Means clustering algorithm was utilized to perform spatial clustering analysis on the time series data, categorizing them into 20 unique clusters within the study area, as depicted in Figure 3. The final clustering result curves illustrate how the cluster centers evolve during the iterative process and how data points progressively stabilize within specific clusters (Figure 4). The clustering results were interpreted by examining the various vegetation conditions and growth states, as detailed in

Table A1. NDVI feature in each classification.

Category	NDVI Feature	Classification
1	Low and relatively stable	Urban areas or water bodies
2–4	High, with seasonal variation	Good vegetation coverage
5	Extremely high in the 10th month	Abnormally high vegetation growth or biomass
6–7	Generally high, exceptionally high in the 10th month	Abundant vegetation or agricultural activities
8–10	High with fluctuations	Natural growth cycles or human activities
11–12	Low, extremely low in the 7th month	Little or no vegetation coverage
13	Stable, no seasonal variation	Relatively consistent vegetation coverage
14–19	Pronounced seasonal changes	Increase during the growing season, decline in autumn and winter
15–19	Significant jump in the 6th and 7th months	Rapid growth or climatic events
18–20	Lowest in the 12th month	Less vegetation coverage in winter

and

Table A2. VH feature in each classification.

Category	VH Feature	Classification
1	Average value −25.5 dB, no seasonal variation	Urban areas
2	Average value −25.1 dB, with seasonal fluctuations	Agricultural irrigation and crop growth
3	Higher average value	Areas with more vegetation or rough soil types
4–6	Pronounced seasonal variation	Vegetation growth cycles
7–9	Long-term increasing trend	Increase in surface moisture content
10–12	Lower average value	Urban or rocky terrain
13–15	Anomaly in the 12th month	Extreme climatic events
16–18	Minor fluctuations	Relatively stable surface conditions
19–20	Significant increase at year-end	Seasonal vegetation changes or wetland moisture variations

.

Table A3. Combined characteristics of different land cover types.

Category	Land Cover Types	Characteristics	Description
1, 10–12	Urban Areas	Low NDVI and VH values	Urban regions with smooth surfaces and sparse vegetation
2–4, 6–7	Agricultural Areas	Higher NDVI values with seasonal VH fluctuations	Farmlands, influenced by irrigation and crop cycles
14–19	Natural Vegetation Areas	Pronounced seasonal NDVI variations and VH fluctuations	Areas with natural vegetation, affected by growth cycles and climate
13–15	Special Event-Affected Areas	Anomalies in VH during certain months	Regions impacted by extreme weather like floods or droughts
16–18	Stable Areas	Minor VH fluctuations	Areas with stable conditions, consistent vegetation or terrain
19–20	Wetland or Seasonally Variable Areas	Significant increase in VH values at year-end	Wetlands or areas with seasonal moisture variations

integrates the results of the NDVI change analysis (Table A1) with those from the VH band analysis (Table A2), which are combined as different land cover types.

Comprehensively considering the land cover classification map and the time series clustering graph of VH and NDVI, we can carry out a more in-depth analysis and understanding.

1. Correlation between land cover type and VH/NDVI value:

Urban areas may show stable low values in the VH chart (around −25.5 dB) because artificial surfaces reflect less radar waves. At the same time, in the NDVI map, these areas may show low values because there is less vegetation cover. Agricultural regions may show seasonal fluctuations in the VH map, reflecting the effects of irrigation and crop growth. In the NDVI map, these areas may show high values during the growing season. Forest areas are likely to show high values in both maps, with higher VH values, because the forest has a strong reflection of radar waves and higher NDVI values because the forest has rich vegetation cover.

2. Seasonal changes:

The time series maps of both VH and NDVI show significant seasonal variations, which echo the color distribution in the land cover classification maps. For example, an increase in VH values in the summer may correspond to the expansion of green areas in the land cover map, which may represent areas that are heavily vegetated in the summer. The peak of NDVI occurs in spring and summer, which is consistent with the vegetation growth cycle shown in the land cover map.

3. Impact of human activities:

Agricultural irrigation and crop growth are clearly reflected in both VH and NDVI maps. Seasonal fluctuations in VH values may match the distribution of agricultural land in the land cover map. The influence of urbanization on VH and NDVI is also obvious. The urban area shows a stable low value in the VH map and a low value in the NDVI map, which is consistent with the distribution of the urban area in the land cover map. These charts and data offer insights into land cover and vegetation changes, aiding potato field identification.

3.2. Potato Identification Model Based on Spatial Clustering and Multi-Indices

Spatial clustering results identified using multi-source time-series remote sensing features and the K-Means clustering algorithm were used in combination with measured potato distribution samples (based on the Section 2.4.2 and Section 3.1) to finally determine the distribution of potatoes in 2021 (Figure 5). The F1 score for the model in monitoring the potato distribution area was 0.95, as detailed in

Table 3. Potato distribution identification model performance using the K-Means algorithm with the time-series data from Sentinel-1/2.

Evaluation Indicators	Value
Accuracy	0.90
Precision	0.90
Recall	1.00
F1 Score	0.95

. This score, being close to 1, signifies that the model was adept at identifying the majority of true-positive samples while keeping the false alarm rate low. The analysis of the experimental outcomes demonstrated that the integration of the K-Means algorithm with the time-series data from Sentinel-1/2 was highly effective for spatial cluster analysis.

Four scenes were selected to illustrate the spatial details of the classification. The first row in Figure 6 shows real Google Earth images, while the second row displays the results of potato classification, with the green area representing the model’s identified distribution area of potatoes and black dots indicating the ground-measured distribution points. Figure 6a,b provides a detailed comparison of the classification results at different spatial resolutions: (a) shows the overall distribution of potato fields at a low spatial resolution, capturing broad patterns; while (b) highlights the model’s ability to identify smaller, more dispersed fields at a high spatial resolution, demonstrating its versatility in both broad and detailed classifications. Figure 6c shows the algorithm’s classification performance in mountainous regions, where errors are primarily due to terrain-induced shadows. Figure 6d displays the algorithm’s classification of areas with towns and water bodies, noting some remaining fragmented areas. Nevertheless, the algorithm effectively distinguishes between water bodies and vegetation near water, despite the similar morphological color characteristics of some farmland and water bodies in RGB images. By comparing with Google Earth images and field survey data, we find that the classification accuracy of the K-Means method is slightly decreased in some edge regions and mixed pixel regions, but the overall effect is small.

3.3. PLB Separability Results Based on Euclidean Distance

In this study, the Euclidean distance between the Sentinel-2 bands and PLB severity was used to quantify the distinguishability of spectral features. The color gradient of the heatmap ranges from white (low) to blue (high), representing the magnitude of the distances (Figure 7). The Euclidean distance matrix shows the distance between different Sentinel-2 bands. A smaller distance indicates that the two bands are more similar in terms of spectral features; a larger distance indicates that they are more different in terms of spectral features. As shown in Figure 7, in general, the spectra were more effective in distinguishing between different PLB severities when there was a large difference in disease severity. Moreover, it was observed that the spectral difference between PLB 85 and other PLB < 85 was particularly large.

3.4. The Characteristics and Importance of the PLB Surveillance Model

Feature importance analysis in Figure 8 shows that the top-ranked features in the model were mainly concentrated in the fourth month (August), which suggests that this period may be a critical stage for disease development, where physiological and morphological features of the vegetation are critical for PLB prediction. In particular, near-infrared bands (e.g., B8_3 and B8A_3) shows consistently high importance throughout the time series, highlighting the key role of vegetation biomass and health in disease prediction. The SAVI_3 and NDWI_3 indices are crucial in the fourth month, highlighting the role of vegetation water status in disease progression. The B2_1 band’s importance in the first month indicates early disease spectral characteristics for early detection. Other significant indicators include B7_3, B6_3, and DWSI3_3, which may relate to vegetation moisture and physiological activity. Vegetation texture features, like B8_savg_3, show the effects of disease on structure. However, features like MSK_CLDPRB, MSK_SNWPRB, and NDBI had zero importance at various times, suggesting they were less relevant for prediction.

3.5. Performance of PLB Surveillance Model

Table 4. Performance of seven PLB surveillance models.

Models	Training RMSE	Validation RMSE	Training R²	Validation R²
CART	25.59	36.42	0.46	0.12
GTB	39.24	42.05	0.63	0.52
RF	16.41	23.40	0.84	0.62
STS–RF	26.23	30.54	0.50	0.10
TS–RF	15.40	27.84	0.81	0.54
MSTS–GTB	16.53	30.50	0.79	0.45
MSTS–RF	14.69	20.50	0.89	0.71

compares seven models: CART, GTB, RF, TS–RF, STS–RF, MSTS–GTB, and MSTS–RF. The comparison results of the three machine learning models show that RF performs best in PLB monitoring and can be used as the main model for PLB monitoring research, followed by GTB. It can be seen from the experimental results that the STS–RF method using single source radar data has poor performance on the verification set, and the verification for R² is only 0.10, indicating that its prediction ability for new data is weak. In contrast, the TS–RF method using single-source optical data performed well on both the training set and the validation set, with a validation R² of 0.54, indicating that it has some predictive power for new data. MSTS–RF, which combines multi-source data with time series, has the best performance on multiple evaluation indexes, the lowest RMSE and the highest R², and has good prediction accuracy and generalization ability. These results suggest that MSTS–RF be prioritized as the final model, as it not only performs well in the training phase but also has the best adaptability to new data in actual deployments. The MSTS–RF model not only has high fitting accuracy to the training data but also has good generalization ability for the unseen validation data. In contrast, the MSTS–GTB model performed well on the training data (RMSE 16.53, R² 0.79), but poorly on the validation data (RMSE 30.50, R² 0.45).

The training set of the MSTS–RF model displayed in Figure 9 had good monitoring performance. The RMSE value was 14.69, which indicated the deviation between the percentage of PLB monitored by the model and the actual observations. The R² value was 0.89, which suggests that the model explains 89% of the variation in the actual observations. Figure 10 shows the performance of the PLB monitoring model on an independent validation set. The validation set provided a more rigorous test of the model’s generalization ability than the training set. The RMSE on the validation set was 20.50, which was higher than the RMSE of 14.69 on the training set. The R²on the validation set was 0.71, which was lower than the R² of 0.89 on the training set.

From the learning curve (Figure 11), we can observe the following: With the increase in the proportion of the training set, the RMSE of the training set decreases gradually, which indicates that the model fits the training data better and better. The RMSE of the validation set is the lowest when the ratio of training set is 0.70, which indicates that the generalization ability of the model is the strongest at this time. When the proportion of the training set decreases further, the RMSE of the validation set increases significantly, indicating that the generalization ability of the model decreases.

The R² of the training set is high at all scales, indicating that the model fits well on the training data. The R² of the validation set reaches the highest when the ratio of the training set is 0.70, indicating that the generalization ability of the model is the strongest at this time. When the proportion of the training set is reduced to 0.10, the R² of the validation set decreases significantly, indicating that the generalization ability of the model decreases significantly.

From these learning curves, we can draw the following conclusions: Optimal data partition ratio: When the training set ratio is 0.70, the model has the best performance on the validation set, the lowest RMSE and the highest R².

Overfitting and underfitting: When the proportion of the training sets is too high or too low, the model is prone to overfitting or underfitting. When the training set ratio is 0.90, the model performs well on the training set, but the performance on the validation set decreases, indicating overfitting. When the training set ratio is 0.10, the performance of the model on the validation set decreases significantly, indicating underfitting.

3.6. The Distribution of PLB Monitoring Results

This figure (Figure 12) shows the distribution of pixel values of PLB severity detected by the surveillance model. The histogram shows that the distribution of PLB values ranges from 4.75 to 68.25, indicating that the monitored PLB severity ranges from very mild to very severe. Most of the pixel values were clustered around 50, indicating that most of the monitored PLB severity was moderate. At the lower (near 4.75) and higher (near 68.25) ends of the PLB values, the number of pixels is lower, indicating that extremely mild or extremely severe PLB conditions are less common. Since most of the predicted values are concentrated in the early and middle stages, this suggests that the monitoring model has the potential to monitor PLB in its early stages, thus helping to take timely measures to control the development of the disease. Overall, this graph shows that the surveillance model is able to better identify moderate severity of PLB and monitor it at early stages, which is important for the management and control of potato late blight.

Figure 13 illustrates the severity of the disease monitored at various locations within the study area, with higher values denoting a greater disease risk as evaluated by the model. The distribution plot highlights substantial spatial heterogeneity in disease risk across the study area. High-risk areas, characterized by monitoring values nearing 80, tend to be clustered in particular geographical locations. This clustering may be associated with local environmental conditions, crop cultivation practices, or historical disease incidence patterns. The disease probability distribution map constitutes a potent tool for agricultural managers, facilitating the identification and prioritization of disease control measures. By targeting these high-risk areas, resources can be more efficiently allocated, thereby enhancing the efficacy and effectiveness of disease management strategies.

Most of the validation points in Figure 14 have small errors, with the majority of errors concentrated in lower intervals (below 25). Points with larger errors tend to be more spatially homogeneous. With an RMSE of 26.52, these results suggest that the model performs well.

4. Discussion

4.1. Sensitivity Features Analysis for PLB

In our study, through feature importance analysis, we identified several key remote sensing metrics that are critical for PLB prediction. Among them, the near-infrared band data (B8_3) in the fourth month (the month of actual data collection and the later period when the disease level develops more severely) topped the list, indicating that its central role in reflecting the physiological status of vegetation is crucial for accurate disease prediction. The NIR spectral band is instrumental in assessing the vitality of vegetation, as it is highly responsive to the biochemical composition of plant life [74,75], especially changes in chlorophyll content. This is closely followed by the Differential Vegetation Index (DVI_3), whose significance score demonstrates its significant impact in indicating changes in vegetation health [76,77]. The DVI is potentially valuable as an indicator of changes in the structure of plant leaves for recognizing early symptoms of disease. Vegetation texture mean (B8_savg_3) and Soil Adjusted Vegetation Index (SAVI_3) for the fourth month also showed high importance scores, emphasizing the role of vegetation structure and cover in disease development. Notably, the May blue band data (B2_1) received a high importance score. In the Random Forest model, feature importance is determined using mean impurity decrease or out-of-bag mean decrease in accuracy [78]. The blue band’s significant importance could stem from its sensitivity to disease indicators [79], such as chlorophyll changes or leaf structure, which are particularly evident in early disease stages. Additionally, the blue band might interact with other features to enhance model predictions. The model’s stochastic nature could also account for variations in the importance scores of different features [80]. In addition, the significance scores of the mid-infrared band data (B7_3 and B6_3) in the fourth month suggest that they may be related to changes in spectral properties caused by physiological activity or disease in the vegetation. The importance scores for the Differential Water Index (DWSI3_3) and the Normalized Differential Water Index (NDWI_3) in the fourth month significantly indicated the influence of vegetation water status on the risk of disease development. These metrics are crucial for monitoring plant water stress and water use efficiency and may reflect plant responses to disease stress. These findings not only aid in identifying critical periods for disease management but also optimize the model structure, thereby enhancing prediction accuracy. Moreover, they provide a scientific basis for long-term monitoring and identification of cyclical disease patterns.

4.2. The PLB Monitoring Model Based on Multidimensional Data

Prior findings have demonstrated the enhanced efficacy of time-series remote sensing data for disease surveillance [81,82,83,84]. Our investigation harnessed the robust computational framework of GEE to synthesize temporal sequences of remote sensing data derived from an array of satellite-based sensors, thereby generating a comprehensive multi-dimensional dataset that encompasses extensive spectral characteristics, disease-related data, and intricate temporal and spatial dynamics. This multidimensional data has 200 bands and contains a large amount of effective information related to PLB. Specifically, the proposed MSTS–RF model improves the R² by about 43% (from 0.46 to 0.89) on the training set and about 59% (from 0.12 to 0.71) on the validation set compared to the CART model. This enhancement demonstrates the significant advantage of time series models in capturing the dynamic properties and patterns of data, especially with time-dependent data. Our methodology for overseeing PLB within the GEE environment is both viable and revelatory, shedding new light on the potential for enhanced synergy between time-series remote sensing data and machine learning algorithms. This approach paves the way for innovative advancements in the realm of agricultural disease monitoring systems.

4.3. The Model Performance for Early PLB Monitoring

The MSTS–RF model developed in this study showcased robust performance, particularly in the training phase, with a lower RMSE and a higher R², indicating its proficiency in modeling the probability distribution of PLB occurrences. However, a slight increase in RMSE and a decrease in R² on the validation set suggest that the model may require further tuning to enhance its generalization capabilities in new, unseen areas. Despite these nuances, the model’s strength in focusing monitoring values within low-risk intervals for a significant proportion of samples underscores its potential for early disease detection. This is especially significant for large-scale agricultural remote sensing, where the subtlety of early disease symptoms can often go undetected [16,42,85]. The model’s ability to perform well across expansive regions is a substantial advancement, providing a foundation for further improvements aimed at bolstering its reliability and precision in real-world applications. Continuous refinement will be essential to adapt the model to diverse environmental conditions and enhance its predictive accuracy on a broader scale.

4.4. Limitations of the Model and Directions for Improvement

The decline in the model’s generalization ability on the validation set indicates that more feature engineering or algorithm adjustments may be needed. Future work may enhance the model’s generalization and robustness by introducing more remote sensing image features, such as feature changes on different time scales, or by combining environmental variables, such as meteorological data [86]. Another limitation of this study is that the model mainly relies on changes in spectral information [87,88], potentially overlooking other factors affecting the occurrence of PLB, such as soil characteristics and crop variety resistance [89]. A possible solution is that of studying the impact of other environmental factors, such as meteorological conditions and soil characteristics, on the occurrence of PLB and integrating them into the model. In addition, the validation of the model is limited to achieving good results in the region selected in the current study, and there are challenges in whether it can perform well in disease monitoring under different geographical and climatic conditions or in a larger range. Future research can further improve the applicability and accuracy of the model by increasing field verification data and validating the model in different regions. In addition, although the current deep learning method has great potential in remote sensing image processing, the current field data are limited, and the driving effect of the deep learning method is especially insufficient. Specifically, deep learning methods usually require a large amount of training data to learn effective feature representations and model parameters [61,90], and our data volume may not cover enough variability and complexity, resulting in less comprehensive features learned by the model, and gradient updates during training may also be unstable, affecting the convergence of the model. Therefore, we choose traditional machine learning methods for research, which are more stable on small data sets and can achieve better performance through feature engineering and model tuning. In the future, as the amount of data increases, we will explore deep learning methods and make more comprehensive comparisons.

5. Conclusions

In this study, we introduced an innovative approach for large-scale monitoring of PLB using time-series data from Sentinel-1/2 satellites on the GEE platform. This method combined the K-Means clustering algorithm with the RF algorithm, comparing seven distinct algorithmic combinations to predict PLB severity and visualize the spatial distribution of affected areas at a resolution of 10 m. Notably, the proposed MSTS–RF model, which incorporates phenological information, demonstrated superior performance on the training dataset, with a lower RMSE and a higher R². Even on an independent validation dataset, although the RMSE increased to 20.50 and the R² decreased to 0.71, the RF algorithm maintained strong generalization, achieving an F1 score of 0.95 for mapping potato distribution. The generated PLB distribution map provided valuable insights into spatial disease risk, supporting targeted agricultural management. However, the study’s limitations include the need for further feature engineering and validation in diverse regions. Future work will focus on incorporating additional environmental data and exploring deep learning methods to enhance model robustness and accuracy. Overall, this study demonstrated a practical and effective approach for large-scale PLB monitoring, contributing to sustainable agricultural management and disease control.