Deep learning-based classification of wet direct seeded rice, broadcasted direct seeded rice and transplanted rice using drone imagery for precision agriculture

1 Introduction

Advancements in remote sensing technologies have significantly enhanced the monitoring and implementation of precision agriculture (PA) practices aimed at promoting sustainable crop production. PA involves several key stages, including data collection, data analysis, decision making and field management. Among various data acquisition methods, Unmanned Aerial Vehicles (UAV) and drone-based imaging has emerged as a powerful tool due to its high spatial resolution and flexibility. These drone-acquired images are processed to extract information about the objects within the field, often using advanced image processing techniques (Franco-Navarro et al., 2025).

Image analysis methods for drone-acquired data can be broadly categorized into pixel-based and patch/scene-based approaches for image classification within agricultural imagery. In pixel-based classifications, each pixel is analyzed individually and a class label is assigned to it. The aggregation of the labelled pixels provides an overall understanding of the objects present in the image. Patch-based methods analyze a small neighborhood or group of pixels (a patch), assigning a single class label to the entire patch. In contrast, scene-based classification assigns a class label to the entire image, assuming it contains only one dominant object class (Cheng et al., 2020).

The use of drone imagery in agricultural applications has gained momentum in recent years. Bouguettaya et al. (2022), provided a comprehensive review of deep learning (DL) models for classifying UAV images in agricultural settings, highlighting the use of advanced deep learning (DL) approaches such as convolutional neural networks (CNNs) for crop detection. Zualkernan et al. (2023), surveyed CNN based crop classification approaches across different agricultural stages, including pre-harvest, harvest and post-harvest. Treboux and Genoud (2018) provided a stage wise overview of pixel- and patch-based approaches for crop classification using ML models. Chin et al. (2023), explored the applications of CNN for detecting plant disease from drone imagery. Pathak et al. (2023), employed image processing techniques along with ML models such as k-nearest neighbor (KNN), support vector machine (SVM), and random forest (RF) to classify weeds in cornfields. Chew et al. (2020) utilized UAV-acquired RGB imagery in Rwanda to develop a deep learning algorithm/pipeline using CNNs and transfer learning (VGG16 and ImageNet) to identify crops such as banana, maize and legumes.

In addition, Torres-Sanchez et al. (2021), integrated UAV data and neural networks (NN) to detect weeds in cropping systems, achieving an overall accuracy of 80%. Ubaidillah et al. (2022), compared RF, NN and Naïve Bayes classifiers to detect corn diseases, with NN outperforming the other models with 74.44% accuracy. Grijalva et al. (2023) and Yadav et al. (2022) demonstrated the applicability of YOLOv5-based deep learning frameworks in agricultural image analysis and crop monitoring tasks. Recently, Jain et al. (2024) introduced a high-throughput plant phenotyping system that integrates Tiny-YOLOv4 with Chan-Vese segmentation for accurate plant detection and segmentation from RGB photos, showcasing the potential of cutting-edge computer vision methods for agricultural image analysis.

Several studies have also explored different image analysis strategies for crop classification. While conventional pixel-based methods classify individual pixels independently, object-based approaches incorporate spatial and contextual information by analysing groups of pixels. For example, Dwivedi et al. (2022), proposed an object-based image analysis framework for crop detection using both drone and satellite imagery.

Despite these developments, previous studies focusing on rice classification have predominantly relied on pixel-based approaches using satellite and SAR imagery. Rotairo et al. (2019) used MODIS (250 m) and Sentinel-2 (10 m) imagery to classify DSR fields using pixel-level techniques. Fikriyah et al. (2019) used Sentinel-1 C-band SAR imagery with a spatial resolution of 20 m to identify and estimate DSR areas. Gumma et al. (2015) employed a similar approach using MODIS (250 m), Landsat-8 (30 m), and RISAT-1 (18 m) imagery to map DSR areas in Raichur district, Karnataka, India. Although these studies demonstrated the potential of remote sensing for rice monitoring, the spatial resolutions of 10–250 m are substantially coarser than the 1.1 cm/pixel ground sampling distance (GSD) achieved by the UAV imagery used in the present study. Consequently, such imagery may have limited ability to capture fine-scale spatial and structural differences among rice establishment methods.

Rice is commonly cultivated using three establishment methods: (i) Broadcasted Direct Seeded Rice (DSR) (ii) Wet Direct Seeded Rice (Wet DSR) and (iii) Transplanted rice (TR) based on the type of sowing. These methods exhibit distinct spatial and structural characteristics that can be captured using high-resolution UAV imagery. Accurate discrimination among these rice establishment methods is essential for crop mapping, resource optimization, and precision agricultural management. However, studies employing UAV-based DL approaches to distinguish these establishment methods remain limited.

With the recent availability of high-resolution data through UAVs such as drones, there is a clear opportunity to apply advanced DL models for more accurate classification of rice establishment methods. In this study, RGB drone imagery with a GSD of 1.1 cm/pixel was acquired during the Kharif season 2023 from an experimental field located in Bapatla District, Andhra Pradesh, India. The study included three rice establishment methods, namely, DSR, Wet DSR, and TR, each covering approximately 2 acres. The performance of advanced DL models, including CNNs, was evaluated for classifying these three rice establishment methods using metrics such as overall accuracy (OA), precision (P), recall (R), Kappa coefficient (K), F1-score (F1) and mean average precision (mAP).

The objectives of the present study are as follows:

• To develop a UAV RGB image-based framework for patch-level classification of three rice establishment methods, namely, Broadcasted Direct Seeded Rice (DSR), Wet Direct Seeded Rice (Wet DSR), and Transplanted Rice (TR).
• To generate a labelled patch-level image dataset from UAV orthomosaic imagery using GPS-guided plot delineation and a systematic image patch extraction workflow.
• To evaluate and compare the performance of ResNet-50-based hybrid models and YOLO-based deep learning models for classification of rice establishment image patches under field conditions.
• To assess the robustness and consistency of the evaluated models across multiple temporal acquisition stages using standard classification metrics.

2 Materials and methods

2.1 Proposed model for the classification of rice cultivation in UAV images

The functional block diagram of the proposed ResNet-50-based hybrid classification framework for classifying rice cultivation types in UAV images is illustrated in Figure 1. The model begins with the acquisition of high-resolution RGB (Red, Green, Blue) images of agricultural fields using an airborne drone. In the feature extraction stage, the ResNet-50 deep learning architecture is employed to extract key informative features from these images. These features are then passed to a classification module, which assigns one of the three rice cultivation classes: TR, DSR or Wet DSR.

2.1.1 Input image

The input image to the model is characterized by three dimensions: (i) no. Of rows, (ii) no. Of columns and (iii) no. of spectral bands. Each pixel in the image encodes the cumulative spectral reflectance from the ground at a specific spatial location. The rows and columns define the spatial layout, while the spectral bands correspond to reflectance values at different wavelength of the electromagnetic spectrum (in this case, RGB).

2.1.2 ResNet-50 feature extractor

The ResNet-50 model (He et al., 2016) comprises 50 layers of convolution and pooling operations designed to extract hierarchical and discriminative features from input images. The model converts an image into a one-dimensional (1-D) feature vector that captures spatial and spectral characteristics relevant for classification. In the subsequent stage, this informative feature vector is passed to a machine learning classifier for class prediction. The internal architecture of the ResNet-50 feature extractor, consisting of convolution and pooling layers, is illustrated in Figure 2. An input image of size 256 × 256 × 3 is progressively downscaled through convolution and max-pooling operations to extract higher-level feature representations. The extracted 1-D feature vector is subsequently used as input for the machine learning classifiers in the final classification stage.

ResNet-50 was selected as the backbone feature extraction network considering its strong transfer learning capability, stable convergence behaviour, moderate computational complexity, and suitability for relatively smaller UAV image datasets. Since the primary objective of the study was to evaluate the performance of different machine learning classifiers using deep features extracted from UAV imagery, ResNet-50 provided an effective balance between feature representation capability and computational efficiency while reducing the risk of overfitting under limited training data conditions.

2.1.3 Classifier in the final layer

The final classification layer uses the feature vectors extracted from ResNet-50 to predict the rice cultivation type. We evaluated six popular machine learning classifiers at this stage: (i) K-Nearest Neighbor (KNN), (ii) Support Vector Machine (SVM), (iii) Neural Network (NN), (iv) Decision Tree (DT), (v) Random Forest (RF) and (vi) Naïve Bayes (NB).

2.1.3.1 k-nearest neighbor (KNN)

KNN is a non-parametric, instance based supervised learning method originally developed by Fix and Hodges (1951), and is a simple yet popular supervised learning method for classification (Pathak et al., 2023). It is considered a lazy or instance-based learning method since it does not involve specialized training but instead relies on labeled data. KNN classifies a test sample by identifying its k nearest labelled data points based on a chosen distance metric, such as Euclidean distance, and assigns the majority class among them (Babbar et al., 2023). The k value determines the number of neighbors considered, where a smaller k makes the model sensitive to noise, while a larger k smooths decision boundaries but may overlook local patterns.

2.1.3.2 Support vector machine (SVM)

SVM was originally developed by Vapnik (1964), is a supervised learning algorithm that finds an optimal hyperplane in N-dimensional space to separate classes while maximizing the margin and minimizing misclassifications. It is particularly effective when combined with kernel functions that transform non-linear decision boundaries into linear ones in a higher-dimensional space (Nodirbek o’g‘li, 2025). SVM has shown strong performance in agricultural classification tasks (Al-Naeem et al., 2023).

2.1.3.3 Neural network (NN)

Artificial Neural Network (ANN) or NN are adaptive, data-driven, non-linear and non-parametric statistical models inspired by biological neural system and are used for tasks ranging from classification to estimation and detection. They consist of input, hidden, and output layers. Unlike traditional models, NNs are capable of learning complex, non-linear relationships between inputs and outputs without requiring explicit feature engineering (Frimpong et al., 2025; Jueyendah et al., 2025). They are widely used in image classification tasks in agriculture.

2.1.3.4 Decision tree (DT)

DT are non-parametric classification method that doesn’t rely on statistical assumptions about data distribution. They consist of a hierarchial structure of nodes, where internal nodes represent decision criteria, branches represent decision paths, and leaf nodes represent final classification (Breiman, 2001). They are intuitive and widely applied in agricultural classification problems (Silva et al., 2025).

2.1.3.5 Random forest (RF)

RF algorithm introduced by Breiman, 2001 is a popular, widely used non-parametric ensemble method for classification and prediction tasks (Hasan et al., 2023). It constructs decision trees during training and combines their outputs using majority voting scheme for classification (Syed, 2024). RF is particularly effective in handling large, complex datasets and high-dimensional feature spaces and is also useful in estimating feature importance. By randomly selecting variable subsets of features at each split, the algorithm reduces correlation between trees and improves overall model stability and accuracy (Salman et al., 2024). The performance of an RF model depends on key hyperparameters such as the number of trees in the forest and the number of features considered at each split (Savitha and Talari, 2025).

2.1.3.6 Naïve bayes (NB)

NB algorithm is a simple yet effective supervised machine learning method based on probabilistic principles (Raja et al., 2022). It assumes conditional independence among features, meaning the presence or absence of a particular feature is considered unrelated to the presence of any other feature within the same class. Despite this strong assumption, NB often performs well in various classification tasks due to its efficiency and robustness, especially with high dimensional data. NB is derived from Bayes’ theorem, which describes the probability of a hypothesis A given the evidence B, and is mathematically expressed as

P A | B = P B | A P A P B

Where, P (A/B) is the posterior probability of class A given the observed feature B, P(B|A) is the likelihood of observing feature B given class A, P(A) is the prior probability of class A before observing the evidence and P(B) is the marginal probability of observing feature B.

2.1.4 Class label

In the testing stage, the model assigns a class label to each input image, indicating the detected rice cultivation method (Transplanted, DSR, or Wet DSR). These labels represent the final output of the model and indicate the rice establishment category represented in the image patch.

2.2 Study area and dataset description

2.2.1 Study area

The study was conducted during the Kharif season 2023 in the experimental field of Praanadhaara Organised Agro Forestry Private Ltd. Located in Cheruvu Jammulapalem, Bapatla District, Andhra Pradesh, India (15.936,502°N 80.440,997°E). The region experiences a tropical climate, characterized by hot summers, mild winters and significant monsoon rainfall, which makes it well suited for rice cultivation.

In recent years, DSR has been widely adopted across the area, driven by support from Praanadhaara Organised Agro Forestry Pvt. Ltd., which plays a key role in promoting modern agricultural practices, including customized mechanization tailored to local farming needs. The geographical location of the study area is shown in Figure 3.

2.2.2 Equipment’s used in this study

A GPS device (Garmin GPSMAP 66SR, United States) was employed to collect precise geolocation data, including latitude, longitude, and elevation, of the experimental plot. The recorded observations were stored as point GPX files and later exported to ESRI shapefile format using GIS software (QGIS version 3.32.1-Lima) for further spatial analysis.

A drone (DJI Mavic 2 Pro, China) (Figure 4a), equipped with a Hasselblad camera featuring a 1-inch CMOS sensor and a 20-megapixel imaging system, was used for RGB aerial image acquisition. For UAV operation and flight monitoring, a drone controller paired with a smartphone was used (Figure 4b). The controller was configured in waypoint mode to enable automated flight operations along predefined flight paths, while the smartphone interface displayed the flight route and waypoints, thereby ensuring precise and consistent UAV navigation during image acquisition.

2.2.3 Image acquisition

Three distinct establishment methods were implemented for cultivating rice for the study: TR, DSR, and wet DSR, each covering approximately 2 acres of land. Spectral images of these plots were acquired using a drone equipped with RGB camera (see previous section) that records reflectance in the Blue (450–495 nm), Green (495–570 nm), and Red (620–750 nm) bands of the electromagnetic spectrum.

Each spectral band provides specific insights into the agricultural landscape: The red band is essential for evaluating vegetation health due to its sensitivity to chlorophyll absorption. The green band enhances the contrast of vegetation and water bodies, improving visual interpretation. The blue band is useful for distinguishing between land and water surfaces and contributes to atmospheric correction. By combining these three spectral bands, full true colour composite images were generated that closely resemble natural human vision, aiding in visual feature identification and classification.

Image acquisition was performed at three different time points during crop development: 15 October 2023 (T₁), 27 October 2023 (T₂), and 01 December 2023 (T₃). DSR and Wet DSR were sown on 03 October 2023, whereas TR seedlings were transplanted on 03 November 2023 following nursery raising. Consequently, the T₁ and T₂ acquisitions corresponded to the early and mid-establishment stages of DSR and Wet DSR and to the pre-transplant field preparation stage of TR, whereas the T₃ acquisition corresponded to the late vegetative stage after transplanting for TR. The UAV was operated in automated waypoint mode at a constant altitude of 40 m above ground level, with 80% front overlap and 75% side overlap to maintain uniform scene coverage and photogrammetric consistency. Flights were carried out between 10:00 and 11:30 a.m. under clear-sky conditions, with wind speeds 3–4 m/s to ensure platform stability and minimize illumination variability. The drone maintained a flight speed of 5.0 m/s, producing approximately 425 raw RGB images per mission (including the buffer area), corresponding to a ground sampling distance (GSD) of 1.1 cm/pixel.

The acquired raw RGB images were subsequently processed into georeferenced orthorectified TIFF (GeoTIFF) formats by Vyomik Innovations Private Limited, a professional drone service provider, using standard photogrammetric workflows. The processed orthomosaic datasets were then used for subsequent image subsetting, feature extraction, and model development. These standardized acquisition and processing protocols, together with controlled crop establishment timelines, ensured that the dataset consistently captured the spatial, spectral, and textural characteristics of each rice establishment method while minimizing environmental and operational variability. Representative UAV images from the three acquisition dates are shown in Figures 5a–c.

2.2.4 Dataset generation and preparation workflow

Figure 6 illustrates the overall workflow adopted for UAV image acquisition, dataset generation, preprocessing, patch extraction, and classification in the present study. The dataset generation process involved GPS-based delineation of experimental plots followed by generation of plot boundary shapefiles using QGIS software. UAV RGB images were acquired using a DJI Mavic 2 Pro drone during three temporal stages, namely, T₁, T₂, and T₃. The acquired UAV images were orthorectified and converted into GeoTIFF orthomosaic datasets for further processing. The plot shapefiles were subsequently overlaid on the orthomosaic images to extract plot-wise regions of interest (ROI) corresponding to DSR, Wet DSR, and TR treatments. The extracted ROI images were further subdivided into 400 × 400 pixel image patches to generate the labelled image dataset. A total of 1,966 labelled image patches were generated, comprising 672 DSR images, 612 TR images, and 682 Wet DSR images. The dataset was partitioned into training and testing subsets and subsequently used for classification using ResNet-50-based machine learning classifiers and YOLO-based deep learning models.

To ensure that the classification model learned differences attributable specifically to rice establishment methods, all three systems, namely, DSR, Wet DSR, and TR, were managed under identical environmental and agronomic conditions. DSR and Wet DSR were sown on 03 October 2023, while TR seedlings were transplanted on 03 November 2023 following nursery raising. Drone image acquisitions were conducted on 15 October 2023 (T1) and 27 October 2023 (T2), corresponding to the early and mid-establishment stages for DSR and Wet DSR, and pre-transplant field preparation stages for TR. A third UAV flight was conducted on 01 December 2023 (T3), representing the late vegetative stage after transplanting for TR. However, since TR transplantation was performed on 03 November 2023, the UAV images acquired during T1 (15 October 2023) and T2 (27 October 2023) primarily represented pre-transplant field conditions for the TR treatment rather than fully developed transplanted rice canopy characteristics. Therefore, the classification performance observed during these early temporal stages should be interpreted mainly as discrimination of establishment-related field and plot-level conditions. These early temporal stages mainly captured differences related to land preparation, soil background, water status, and pre-transplant field conditions, whereas the T3 imagery captured more representative crop canopy characteristics across all three rice establishment methods.

2.2.4.1 Image sub-setting

Sub-setting an image is the process of extracting a specific region of interest (ROI) from a large orthomosaic image based on pre-defined spatial boundaries. In this study, the experimental plot boundaries were created as shapefiles by digitizing GPS observations taken with a Garmin GPS device and processed in QGIS software. The created shapefiles were overlaid on the UAV RGB orthomosaic images to demarcate the experimental plots for Transplanted Rice (TR), Broadcasted Direct Seeded Rice (DSR), and Wet Direct Seeded Rice (Wet DSR) as shown in Figure 5. The delineated regions of interest were extracted and split into 1966 image patches of size 400 × 400 pixels using Google Colab for model development and evaluation.

2.2.4.2 Dataset

The UAV images of the experimental plots were subdivided into 1,966 sub-images, which were used to prepare the dataset for model development and evaluation. A summary of the dataset composition is provided in Table 1. The sample images of DSR, TR, and Wet DSR are shown in Figures 7a–c, respectively. The images in the dataset were divided into training and testing subsets. The training set images were used to train the models, while the testing set images were used to evaluate model performance.

TABLE 1. Details of the image dataset.

2.3 Model description

In this study, six CNN based hybrid models were developed for classifying rice cultivation types. Each model consists of two core components: a feature extractor and a classifier. The ResNet-50 architecture was employed as the feature extractor across all models, while different machine learning classifiers were used in the final classification stage. ResNet-50 was selected as the backbone feature extraction network considering its strong transfer learning capability, computational efficiency, and suitability for relatively smaller UAV image datasets. Since the present study involved a moderate number of training images, the use of substantially deeper architectures could increase the risk of overfitting. Therefore, ResNet-50 was considered an appropriate balance between model complexity and generalization performance for extracting spatial and textural features associated with different rice establishment methods.

The following six models were constructed:

• Model 1: ResNet-50 + DT
• Model 2: ResNet-50 + RF
• Model 3: ResNet-50 + SVM
• Model 4: ResNet-50 + NB
• Model 5: ResNet-50 + KNN
• Model 6 (proposed): ResNet-50 + NN

These models leverage the high-level feature representations extracted by ResNet-50 and apply different classification algorithms to label the images into one of the three rice establishment types: Transplanted Rice (TR), Broadcasted Direct Seeded Rice (DSR), or Wet DSR.

2.3.1 YOLO-based deep learning models

In addition to ResNet-50 hybrid models, two other deep learning architectures, namely, You Only Look Once version 5 (YOLOv5) and You Only Look Once version 8 (YOLOv8) were also used for classification in this study. These models were selected because of their widespread adoption, computational efficiency, and proven effectiveness in agricultural image analysis, crop monitoring, and object recognition applications. Furthermore, YOLOv8 represents a recent advancement over YOLOv5, offering improvements in feature extraction, learning capability, and computational efficiency. The inclusion of both models enabled a comparative evaluation of an established YOLO architecture and its newer-generation counterpart for classifying rice establishment methods using UAV imagery.

YOLOv5 is a deep learning framework that can also be adapted for image classification tasks using learned spatial feature representations. It consists of four main components: input, backbone, neck, and head. The backbone uses Spatial Pyramid Pooling (SPP) and Cross-Stage Partial network (CSP) for feature extraction (Wang et al., 2019). The neck employs PANet (Path Aggregation Network) to enhance multi-scale feature fusion. The head is responsible for generating the final predictions.

YOLOv8, one of the recent models in the YOLO series, offers improvements in computational efficiency and feature learning capability for image analysis applications. Similar to YOLOv5, it incorporates an optimized backbone, neck, and head, and supports advanced functionalities like instance segmentation. YOLOv8 utilizes an enhanced version of the Darknet-53 backbone (Wang et al., 2022), and introduces improved feature learning and prediction capabilities for image classification across varying spatial patterns.

2.4 Performance metrics

In the present study, the performance of the classification models was evaluated using the following metrics: OA, P, R, F1, K and mAP. OA represents the proportion of correctly classified samples out of the total number of samples. In this study, OA was computed as the average accuracy across different proportions of training data: 10%, 20%, 30%, …., up to 90%. Precision is defined as the ratio of the true positive predictions to the total number of predicted positives (sum of true positives and false positives). Recall is known as sensitivity and is the ratio of the number of correctly predicted positives divided by the total number of predicted positives. F1-score is the harmonic mean of P and R. F1-score lies between 0 and 1. The model with F1 score near to 1 indicates better performance. The Kappa coefficient measures the agreement between predicted and actual classifications, adjusted for the agreement occurring by chance. It complements OA and is particularly useful for imbalanced datasets. mAP was used as an additional evaluation metric to assess model discrimination performance across rice establishment classes. Higher mAP values indicate improved class-level prediction performance and model robustness. In general, the higher the value of OA, K, P, R, F1 and mAP indicates the better model performance.

3 Results

3.1 Feature extraction using ResNet-50

To extract features using ResNet-50, the pre-trained model was employed to extract high-level features from the input images. The output consisted of a 2048-dimensional feature vector for each image, which was subsequently stored in a CSV file. Each row of the file corresponded to an individual image and its associated feature vector, while the final column contained the corresponding class label. As detailed in Section 2.1.2, the extracted ResNet-50 features were subsequently used as inputs to the various machine learning classifiers, resulting in the hybrid models ResNet-50 + DT, ResNet-50 + RF, ResNet-50 + SVM, ResNet-50 + NB, ResNet-50 + KNN, and ResNet-50 + NN.

3.2 Dataset partitioning and model training

The dataset comprising 1,966 labelled sub images was divided into two disjoint training and testing subsets. To evaluate model performance under different training data conditions, the training set was incrementally varied across 10%, 20%, 30%, 40%, 50%, 60%, 70%, 80%, and 90% of the fully labelled images dataset. Model parameters were optimized using the training set, and classification accuracy was assessed on the test set. The detailed hyperparameter configurations of all models are provided in the supplementary material (Supplementary Tables S1–S7). All models were trained and evaluated on a system with the following configuration: Intel(R) Core(TM) i7-2600, CPU @ 3.40GHz, 4 Cores, 8 Logical Processors.

3.3 Performance analysis

The confusion matrices for various models across the three time intervals (T₁, T₂, and T₃) are summarized in Supplementary Table S8, while the corresponding graphical confusion matrices are presented in Supplementary Figures S1–S3. These matrices serve as the foundation for calculating key performance metrics, including OA, P, R, F1, and K. This detailed evaluation overview provides a comprehensive comparison of each classifier’s strengths and limitations in identifying the three rice establishment methods.

For instance, Supplementary Figures S1a shows the confusion matrix for the ResNet-50+DT classifier for T₁ interval. The data was split into 60 per cent for training and 40 per cent for testing. A total of 786 images were used for testing. Out of these, 280 images were classified into the DSR (Class-0). Among 280 images, 204 images were correctly classified as DSR (Class-0), 48 were incorrectly classified as TR (Class-1) and 28 were misclassified as Wet DSR (Class-2). Out of 228 images of TR, 175 images were correctly classified, 40 were misclassified as DSR and 13 were misclassified as Wet DSR. Out of 278 images of Wet DSR, 239 were correctly classified, 33 were wrongly classified as DSR and 6 were misclassified as TR.

3.3.1 Model accuracy across time-intervals

The classification accuracy results of all evaluated models under various train–test splits (10:90 to 90:10) at T₁, T₂, and T₃ are summarized in Tables 2–9, Tables 10–17, and Tables 18–25, respectively. At time interval T₁, the ResNet-50+NN model achieved the highest average overall accuracy (93.16%), followed by ResNet-50+SVM (92.07%), ResNet-50+KNN (89.08%), ResNet-50+RF (87.69%), YOLOv8 (80.27%), ResNet-50+DT (78.78%), YOLOv5 (78.37%), and ResNet-50+NB (78.05%). Similarly, at time interval T₂, ResNet-50+NN again recorded the highest average overall accuracy (95.81%), followed by ResNet-50+SVM (95.24%), ResNet-50+RF (93.65%), ResNet-50+KNN (93.51%), YOLOv8 (88.94%), YOLOv5 (88.60%), ResNet-50+NB (88.39%), and ResNet-50+DT (86.11%). At time interval T₃ (Tables 26–28), ResNet-50+NN remained the best-performing model with the highest average overall accuracy (93.31%), followed by ResNet-50+SVM (92.09%), ResNet-50+KNN (90.09%), ResNet-50+RF (88.55%), YOLOv8 (83.93%), YOLOv5 (81.83%), ResNet-50+DT (79.69%), and ResNet-50+NB (77.47%). Overall, the ResNet-50+NN classifier constantly achieved the highest mean accuracy across all three time-intervals, establishing it as the top-performing model among all evaluated classifiers.

TABLE 2. Accuracy assessment of ResNet-DT at time interval T₁.

Bold values represent the average overall accuracy.

TABLE 3. Accuracy assessment of ResNet-RF at time interval T₁.

Bold values represent the average overall accuracy.

TABLE 4. Accuracy assessment of ResNet-SVM at time interval T₁.

Bold values represent the average overall accuracy.

TABLE 5. Accuracy assessment of ResNet-NB at time interval T₁.

Bold values represent the average overall accuracy.

TABLE 6. Accuracy assessment of ResNet-KNN at time interval T₁.

Bold values represent the average overall accuracy.

TABLE 7. Accuracy assessment of ResNet-NN at time interval T₁.

Bold values represent the average overall accuracy.

TABLE 8. Accuracy assessment of YOLOv5 at time interval T₁.

Bold values represent the average overall accuracy.

TABLE 9. Results of the trained YOLOv8 model at time interval T₁.

Bold values represent the average overall accuracy.

TABLE 10. Classification accuracy (%) of models at time interval T₁.

TABLE 11. Accuracy assessment of ResNet-DT classifier at time interval T₂.

Bold values represent the average overall accuracy.

TABLE 12. Accuracy assessment of ResNet-RF at time interval T₂.

Bold values represent the average overall accuracy.

TABLE 13. Accuracy assessment of ResNet-SVM at time interval T₂.

Bold values represent the average overall accuracy.

TABLE 14. Accuracy assessment of ResNet-NB at time interval T₂.

Bold values represent the average overall accuracy.

TABLE 15. Accuracy assessment of ResNet-KNN at time interval T₂.

Bold values represent the average overall accuracy.

TABLE 16. Accuracy assessment of ResNet-NN at time interval T₂.

Bold values represent the average overall accuracy.

TABLE 17. Accuracy assessment of YOLOv5 at time interval T₂.

Bold values represent the average overall accuracy.

TABLE 18. Results of the trained YOLOv8 model at time interval T₂.

Bold values represent the average overall accuracy.

TABLE 19. Classification accuracy (%) of models at time interval T₂.

TABLE 20. Accuracy assessment of ResNet-DT at time interval T₃.

Bold values represent the average overall accuracy.

TABLE 21. Accuracy assessment of ResNet-RF at time interval T₃.

Bold values represent the average overall accuracy.

TABLE 22. Accuracy assessment of ResNet-SVM at time interval T₃.

Bold values represent the average overall accuracy.

TABLE 23. Accuracy assessment of ResNet-NB at time interval T₃.

Bold values represent the average overall accuracy.

TABLE 24. Accuracy assessment of ResNet-KNN at time interval T₃.

Bold values represent the average overall accuracy.

TABLE 25. Accuracy assessment of ResNet-NN at time interval T₃.

Bold values represent the average overall accuracy.

TABLE 26. Accuracy assessment of YOLOv5 at time interval T₃.

Bold values represent the average overall accuracy.

TABLE 27. Results of the trained YOLOv8 model at time interval T₃.

Bold values represent the average overall accuracy.

TABLE 28. Classification accuracy (%) of models at time interval T₃.

3.3.2 Classifier performance metrics

Figures 8a–c illustrate the training and validation performance curves of the YOLOv8 model, including loss, precision, recall, and mean Average Precision (mAP) across training epochs. These epoch-wise curves provide important insights into model convergence behaviour, training stability, and potential overfitting patterns. The progressive reduction in loss values and the corresponding improvement in precision, recall, and mAP indicate stable learning and effective convergence of the YOLOv8 framework throughout the training process. Since YOLOv8 is trained using an end-to-end deep learning approach, such training dynamics are inherently generated and enable a comprehensive evaluation of model performance. In contrast, the ResNet-50 framework was employed as a pretrained fixed feature extractor without fine-tuning of the backbone network. Therefore, epoch-wise training and validation loss curves were not generated for the ResNet-50 component. Instead, the extracted deep features were used as inputs for conventional machine learning classifiers (DT, RF, SVM, NB, KNN, and NN), whose performance and robustness were assessed using multiple train-test split scenarios and standard classification metrics. Tables 29, 30 and Figure 9 provides a summary of additional performance metrics including precision, recall, F1-score, and kappa coefficient for the ResNet-50-based classifiers and YOLOv8. Among the evaluated classifiers, ResNet-50+NN and ResNet-50+SVM models consistently achieved the highest precision, recall, and F1-score values across all three-time intervals, indicating their robustness and reliability in drone image classification. The ResNet-50+KNN and ResNet-50+RF models also showed strong and stable performance, though they were slightly below the top two. In contrast, the ResNet-50+DT and ResNet-50+NB models performed relatively poorly but still showed moderate improvement over time.

TABLE 29. Performance metrics (P, R, F1, Kappa) for ResNet-based models across time points.

TABLE 30. YOLOv8 detection results across time points T₁, T₂, and T₃.

4 Discussion

The present study demonstrates the effective integration of UAV-based remote sensing and deep learning techniques for discriminating three rice establishment methods, namely, Broadcasted Direct Seeded Rice (DSR), Wet Direct Seeded Rice (Wet DSR), and Transplanted Rice (TR). Unlike many previous studies that focused on crop species classification, plant disease detection, weed identification, or rice area mapping using satellite imagery, this study specifically addresses the classification of rice establishment methods using high-resolution RGB drone imagery acquired at three different time intervals. The temporal imagery provided detailed spatial and textural information that facilitated accurate discrimination among the three establishment methods.

To evaluate classification performance, ResNet-50 was used as a deep feature extractor and combined with several classifiers, including Decision Tree (DT), Random Forest (RF), Support Vector Machine (SVM), Naïve Bayes (NB), k-Nearest Neighbors (KNN), and Neural Networks (NN). In addition, two YOLO-based deep learning architectures (YOLOv5 and YOLOv8) were evaluated under identical experimental conditions. This robust comparative framework enabled assessment of both hybrid and end-to-end deep learning methods for classifying rice establishment methods.

Among all evaluated models, the ResNet-50+NN model (Model 6) configuration consistently achieved the highest classification accuracies across all time intervals, with average accuracies of 93.16%, 95.81%, and 93.31% at T1, T2, and T3, respectively. This superior performance can be attributed to the complementary strengths of the ResNet-50 architecture and the neural network classifier. ResNet-50 extracts hierarchical spatial and textural features from UAV imagery using deep convolutional layers, while the neural network classifier learns complex nonlinear relationships among these features. Consequently, the ResNet-50+NN model demonstrated greater capability to distinguish subtle differences among rice establishment methods compared to the other classifiers. The systematic partitioning of the dataset into multiple training and testing subsets facilitated rigorous model validation and ensured consistent performance evaluation across different temporal scenarios.

The methodology adopted in this study aligns well with emerging trends in precision agriculture, where UAVs like drone-based imaging and deep learning models are increasingly utilized for efficient crop monitoring and decision-making. The use of QGIS software and GPS-derived shapefiles for image subsetting ensured spatial accuracy and streamlined the preprocessing workflow, further enhancing the reliability of the classification results.

The effectiveness of CNN-based hybrid models in agricultural remote sensing has been well supported in the literature. For example, Viskovic et al. (2019), Agarwal and Bachan (2023), and Azarmdel et al. (2020) demonstrated the utility of deep learning for crop classification tasks. Similarly, Bouguettaya et al. (2022) and Chew et al. (2020) highlighted the advantages of deep learning and transfer learning in UAV-based crop classification and monitoring. Grijalva et al. (2023), Loddo et al. (2021), and Pathak et al. (2023) further demonstrated the effectiveness of these methods for pest, seed, and weed classification. In rice systems, Solanki (2024) emphasized the utility of machine learning for stage-wise crop classification. These studies collectively support the increasing adoption of deep learning approaches in agricultural image analysis. The superior performance of the ResNet-50+NN model observed in this study aligns with previous findings and further reinforces the effectiveness of deep learning-based feature extraction for agricultural image classification. The proposed framework, which integrates ResNet-50 deep feature extraction with a neural network classifier, represents a hybrid deep learning approach that leverages the strengths of both architectures.

Another important finding is the robustness of the ResNet-50+NN model across multiple image acquisition dates. Despite variations in crop growth stage, field conditions, soil background, and canopy development, the model consistently achieved high classification performance. This indicates that the extracted deep features were sufficiently informative and stable to capture the distinctive spatial and structural characteristics associated with different rice establishment methods.

Overall, the findings of this study contribute to the growing body of evidence supporting AI-driven approaches in precision agriculture. More importantly, the study demonstrates the feasibility of using high-resolution UAV imagery and hybrid deep learning frameworks to discriminate among rice establishment methods, a topic that has received limited attention in previous research. Accurate identification of DSR, Wet DSR, and TR systems can support crop mapping, monitoring, resource management, and decision-making in rice production systems. The results further highlight the advantages of combining powerful convolutional feature extractors like ResNet-50 with advanced classifiers to achieve high classification accuracy and robust performance in remote sensing applications. Future work could explore the integration of multispectral or hyperspectral imagery to capture additional crop traits, potentially enhancing model performance and applicability even further.

One limitation of the present study is that the T1 and T2 UAV image acquisitions were conducted before transplantation of the TR treatment. Consequently, the classification results obtained during these early temporal stages may partly reflect differences in field preparation, soil and water background conditions, and establishment-related management practices rather than fully developed transplanted rice canopy characteristics. Therefore, caution should be exercised when interpreting early-stage classification performance across all establishment methods.

Another limitation is that DSR and Wet DSR plots were planted on the same date and therefore had similar crop ages during UAV image acquisition. This experimental design was intentionally adopted to isolate differences attributable to rice establishment methods while minimizing confounding effects arising from crop age and phenological variation. However, in real-world agricultural conditions, particularly at regional scales, sowing and transplanting dates often vary substantially among fields due to variations in rainfall patterns, water availability, labor availability, and management practices. As a result, crop age and growth stage can become important sources of variability that may impact classification performance. Therefore, the results reported in this study should be interpreted in the context of synchronized crop establishment conditions. Future studies should evaluate the proposed framework across various planting dates, seasons, and regions to evaluate its effectiveness and applicability under more diverse agricultural environments.

Looking ahead, integrating next-generation artificial intelligence methodologies has significant potential to enhance rice establishment classification and precision crop monitoring workflows. Emerging AI paradigms, including transformer-based vision architectures, self-supervised representation learning, and multimodal fusion models, offer promising ways to capture richer spatial and temporal patterns from UAV imagery while reducing dependence on large labelled datasets. Incorporating explainable AI techniques can improve model transparency and help stakeholders interpret the spectral or textural features that influence classification outputs, thereby supporting informed decision-making. The deployment of lightweight deep learning models on edge devices could enable real-time data processing directly from UAV platforms. Furthermore, integrating multispectral, hyperspectral, and thermal UAV imagery with advanced AI pipelines may facilitate the detection of subtle crop physiological traits related to water status, nutrient dynamics, and early stress responses. These advancements would significantly expand the applicability of the current study and contribute to the development of more resilient, data-driven, and scalable precision agriculture systems.

5 Conclusion

This study investigated the use of advanced deep learning models for classifying drone imagery of three distinct rice establishment methods. High-resolution RGB images, captured at multiple time-intervals using UAVs, provided a robust dataset for training and evaluating model performance. Among the tested models, the hybrid ResNet-50+NN consistently achieved the highest classification accuracy across all time periods. This model outperformed other classifiers, including ResNet-50+KNN, ResNet-50+SVM, ResNet-50+DT, ResNet-50+RF, ResNet-50+NB, as well as deep learning models such as YOLOv5 and YOLOv8. The superior performance of ResNet-50+NN underscores its effectiveness for patch-level classification tasks in agricultural remote sensing under the evaluated temporal conditions, offering significant potential for accurate crop classification and monitoring. These findings contribute to the growing body of research supporting the application of AI in precision agriculture, enabling more informed decision making towards sustainable farming practices. Future research should consider incorporating additional spectral bands (e.g., NIR), multi-temporal data analysis, and real-time model deployment to enhance classification performance and scalability for large-scale agricultural monitoring systems.

Generative AI statement

The author(s) declared that generative AI was not used in the creation of this manuscript.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher’s note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Supplementary material

The Supplementary Material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/frsen.2026.1698781/full#supplementary-material

Supplementary material Supplementary material at the source

Footnotes

Edited by: Jean-Pierre Wigneron, Institut National de Recherche Pour l’Agriculture, l’Alimentation et l’Environnement (INRAE), France

Reviewed by: Juan De Dios Franco-Navarro, Spanish National Research Council (CSIC), Spain

Trong Hieu Luu, Can Tho University, Vietnam