PLANT DISEASE DETECTION USING DEEP LEARNING AND ANDROID-BASED MOBILE APPLICATION

International Journal of Computer Application ISSN 2250-1797 Available online on https://rspublication.com/ijca/ijca_index.htm Volume 15 Number. 6, 2025 DOI: 10.5281/zenodo.17761867 Original Article ©2025 RS Publication, [email protected] 80 PLANT DISEASE DETECTION USING DEEP LEARNING AND ANDROID-BASED MOBILE APPLICATION Swetha B Assistant Professor, Department of Computer Science and Engineering, Dayananda Sagar Academy of Technology and Management, Bengaluru, India Mahant Singh Department of Computer Science and Engineering, Dayananda Sagar Academy of Technology and Management, Bengaluru, India Lakshmi M Department of Computer Science and Engineering, Dayananda Sagar Academy of Technology and Management, Bengaluru, India Manasa M Department of Computer Science and Engineering, Dayananda Sagar Academy of Technology and Management, Bengaluru, India Hamsa M Department of Computer Science and Engineering, Dayananda Sagar Academy of Technology and Management, Bengaluru, India International Journal of Computer Application https://rspublication.com/ijca/ijca_index.htm ISSN 2250-1797 ARTICLE INFO ABSTRACT ©2025 RS Publication Paper ID: IJCA6929778A54187 Published: 2025-11-29 DOI: https://dx.doi.org/ 10.5281/zenodo.1775 9149 Page No: 80-88 Agriculture plays a vital role in sustaining global food supply, and plant diseases significantly affect crop productivity, quality, and economic stability. Conventional disease diagnosis methods rely heavily on manual inspection by experts, which can be time-consuming, inconsistent, and inaccessible to rural farming communities. In recent years, advances in deep learning have enabled automated and highly accurate plant disease classification using leaf images. This research proposes a Convolutional Neural Network (CNN)-based plant disease detection system integrated into an Android mobile application. The system is designed to capture plant images through the device camera or import them from the gallery, process the data using a pre-trained CNN model optimized with TensorFlow Lite, and generate real-time disease predictions. The solution aims to assist farmers in early disease identification, improve decision-making, and reduce crop loss. Experimental evaluation demonstrates high accuracy in disease classification across multiple plant species. The mobile application interface is designed with simplicity, enabling seamless use even by non-technical users. This research contributes an implementable, accessible, and portable plant disease detection tool with potential scalability for real-world agricultural deployment. keywords: Plant disease detection, Convolutional Neural Networks, Mobile application, TensorFlow Lite, Image classification, Deep learning. Cite This Paper: Mahant Singh, Swetha B, Lakshmi M, Manasa M and Hamsa M (2025). "PLANT DISEASE DETECTION USING DEEP LEARNING AND ANDROID-BASED MOBILE APPLICATION". INTERNATIONAL JOURNAL OF COMPUTER APPLICATION (IJCA), vol. 15, no. 6, 2025, pp. 80-88. DOI: https://dx.doi.org/10.5281/zenodo.17759149 International Journal of Computer Application ISSN 2250-1797 Available online on https://rspublication.com/ijca/ijca_index.htm Volume 15 Number. 6, 2025 DOI: 10.5281/zenodo.17761867 Original Article ©2025 RS Publication, [email protected] 81 1. INTRODUCTION Agriculture serves as the backbone of many developing and developed economies. Sustaining crop quality and yield is essential in maintaining food security and ensuring economic balance. Plant diseases are among the major threats to agricultural productivity, causing substantial global crop losses each year. According to recent estimates, plant pathogens are responsible for up to 20–40% yield reduction in various crops annually. Early detection and timely intervention remain the most effective methods for reducing these losses. However, traditional disease identification methods require expert knowledge, manual inspection in farms, and laboratory analysis—all of which may not be available in remote agricultural regions. With rapid advancements in artificial intelligence and machine learning, particularly convolutional neural networks (CNNs), automated plant disease classification systems have gained attention. CNNs have demonstrated significant performance in image recognition tasks due to their ability to automatically extract spatial features from raw images. Their capability to identify disease patterns such as discoloration, lesions, and texture variations makes them suitable for plant disease classification. In parallel, the widespread availability of smartphones and increasing digitalization in rural communities create opportunities for deploying intelligent agricultural tools. Most farmers now possess Android-based smartphones, which can serve as powerful platforms for implementing real-time disease detection solutions. In this research, a deep learning-based plant disease detection system is developed and integrated into a userfriendly Android application. The system enables farmers to either capture a live image of a plant leaf using the smartphone camera or select an image from the device gallery for disease diagnosis. The trained CNN model is converted to TensorFlow Lite format to ensure efficient inference on mobile devices without requiring cloud connectivity. This enhances the system's portability and reliability, particularly in rural regions with limited internet availability. The proposed application is intended to reduce dependency on expert consultants, minimize diagnostic delays, and make disease identification accessible to every farmer. The combination of machine learning and mobile technology presents a practical and scalable solution to transform plant health monitoring in modern agriculture. 2. LITERATURE SURVEY Automated plant disease detection has been extensively researched, with deep learning emerging as a breakthrough approach in recent years. Traditional image processing methods relied on handcrafted features such as color, shape, and texture descriptors. However, manual feature extraction often fails to generalize across varying environmental conditions and multiple disease categories. Consequently, machine learning approaches based on convolutional neural networks have gained prominence for their superior classification accuracy. Early studies focused on classical machine learning models such as Support Vector Machines (SVM) and Random Forest classifiers applied to segmented leaf images. While these methods showed promising results in controlled environments, they struggled in real-world agricultural settings due to variations in luminosity, leaf orientation, and background noise [1]. This limitation prompted the adoption of deep learning models that learn hierarchical features directly from pixel-level data. Research on CNN-based plant disease detection accelerated after the introduction of publicly available datasets such as the PlantVillage dataset. Authors in [2] developed a deep CNN model capable of distinguishing between 26 diseases across 14 plant species. Their study demonstrated the robustness of deep networks for agricultural applications. Other researchers explored transfer learning using pre-trained models such as VGG16, ResNet50, and InceptionV3 to enhance accuracy and reduce training time [3]. These architectures yielded classification accuracies exceeding 95% in controlled datasets. Mobile-based plant disease detection has also been reported in several studies. Researchers in [4] introduced an Android application that employed a cloud-based convolutional network to process images of tomato leaves. However, reliance on cloud servers increased latency and made the system unusable in rural areas with limited internet access. To overcome these challenges, recent research shifted toward on-device inference using lightweight models like MobileNet and TensorFlow Lite [5]. These approaches significantly reduced memory usage and enabled offline disease detection on smartphones. Moreover, various studies have highlighted the importance of user-friendly interfaces for practical adoption by farmers [6]. Mobile applications with simple designs, icons, and multilingual support demonstrated higher usability. Some works proposed integrating recommendations such as pesticide suggestions and preventive measures based on classification output [7]. Despite advancements, existing solutions often face challenges such as large model sizes, limited dataset diversity, and inadequate real-world performance. Many studies train models on uniform backgrounds, causing the systems International Journal of Computer Application ISSN 2250-1797 Available online on https://rspublication.com/ijca/ijca_index.htm Volume 15 Number. 6, 2025 DOI: 10.5281/zenodo.17761867 Original Article ©2025 RS Publication, [email protected] 82 to fail in field conditions where leaves exhibit dust, shadows, or irregular lighting. Additionally, only a few research works integrate the entire pipeline—image acquisition, model inference, disease visualization, and mobile app deployment—into a single operational system. This research builds upon the gaps identified in the literature by developing a complete end-to-end solution combining a TensorFlow Lite-optimized CNN model and a fully functional Android application. The model is trained to detect multiple diseases across plant species, and the application provides real-time predictions, making it a practical tool for farmers. 3. METHODOLOGY The The methodology adopted in this research integrates deep learning–based image classification with an Android mobile platform to create a fast, portable, and efficient plant disease detection system. The approach follows a systematic pipeline beginning from dataset acquisition, preprocessing, CNN model development, TensorFlow Lite conversion, and mobile application integration, ensuring that the system performs reliably under real-world agricultural scenarios. Similar multi-phase methodologies have been adopted in several agricultural AI systems to ensure robustness and field adaptability [8]. 3.1 Dataset Collection and Preparation A high-quality dataset is essential for training a robust plant disease classifier. In this project, publicly available plant disease datasets were used, consisting of thousands of leaf images representing multiple crops and disease categories. The dataset includes healthy plant leaves as well as leaves affected by fungal, bacterial, and viral infections. The use of diverse samples ensures that the model can generalize well in practical environments, as noted in related studies on plant pathology classification [9]. All images were resized to a standard dimension of 224×224 pixels to maintain uniformity and reduce computational overhead during training. Data augmentation techniques such as rotation, flipping, brightness adjustment, and zooming were applied to enrich the dataset and prevent overfitting. Multiple research works highlight that augmentation significantly enhances the model’s resilience to environmental variations [10]. 3.2 CNN Model Architecture A custom Convolutional Neural Network was designed for plant disease classification. The CNN architecture consists of convolutional layers for feature extraction, ReLU activation functions, pooling layers for spatial reduction, and fully connected layers for final classification. This architecture is widely used in image-based tasks due to its strong ability to learn hierarchical patterns such as color distortion, disease lesions, and texture anomalies [11]. Batch normalization and dropout layers were incorporated to enhance model stability and reduce overfitting. The final output layer uses softmax activation to generate probability scores for each disease class. 3.3 Model Training The model was trained using the Adam optimizer with categorical cross-entropy loss. The training was conducted over multiple epochs, ensuring the network learns meaningful patterns from the dataset. Validation accuracy was continuously monitored, and early stopping was applied once the model performance stabilized. Studies in deep learning–based agriculture emphasize the importance of such regularization techniques to maintain accuracy while preventing model overtraining [12]. 3.4 Conversion to TensorFlow Lite To deploy the model efficiently on an Android device, the trained TensorFlow model was converted to TensorFlow Lite (.tflite). This format reduces memory consumption and accelerates inference time, enabling real-time predictions even on low-end smartphones. Similar strategies for model optimization have been successfully implemented in mobile disease detection systems and lightweight AI applications [13]. 3.5 Android Application Integration The final phase involved integrating the TFLite model within the Android application. The application was developed using Java in Android Studio, incorporating camera and gallery functionalities for image acquisition. The image is then preprocessed, normalized, resized, and fed to the TFLite interpreter for prediction. Such on-device inference models have proven effective in rural regions lacking reliable internet service, as demonstrated in agricultural mobile analytics tools [14]. International Journal of Computer Application ISSN 2250-1797 Available online on https://rspublication.com/ijca/ijca_index.htm Volume 15 Number. 6, 2025 DOI: 10.5281/zenodo.17761867 Original Article ©2025 RS Publication, [email protected] 83 4. PROPOSED SYSTEM ARCHITECTURE The overall system architecture is divided into four major components: image acquisition, preprocessing, disease classification, and result display. These components work together to provide a seamless and interactive experience for the user. 4.1 Image Acquisition Module Users can acquire images either by capturing a photo through the device camera or selecting an image from the gallery. The system supports multiple Android versions by adopting the Storage Access Framework (SAF) to ensure proper URI handling and access permissions. Prior studies highlight the importance of efficient image acquisition in mobile-based diagnostic systems [15]. 4.2 Preprocessing Module Once the image is acquired, it undergoes preprocessing such as resizing to 224×224 pixels and normalization to scale pixel values between 0 and 1. These steps prepare the image for efficient feature extraction by the CNN. Preprocessing is a crucial step to ensure consistent model performance under varying lighting and environmental conditions [16]. 4.3 Disease Classification Module The TFLite interpreter loads the trained CNN model and runs inference on the preprocessed image. The output is a probability vector representing the confidence score for each disease class. The class with the highest probability is selected as the predicted result. This approach is consistent with standard multi-class classification frameworks used in plant pathology research [17]. 4.4 Result Visualization and User Interface The final stage involves displaying the predicted disease, accuracy percentage, and additional description through a user-friendly interface. The simple layout ensures that farmers and non-technical users can understand the results easily. Previous research emphasizes the need for intuitive interfaces in digital agriculture tools to encourage International Journal of Computer Application ISSN 2250-1797 Available online on https://rspublication.com/ijca/ijca_index.htm Volume 15 Number. 6, 2025 DOI: 10.5281/zenodo.17761867 Original Article ©2025 RS Publication, [email protected] 84 adoption by rural communities [18]. 5. IMPLEMENTATION The proposed system was implemented using two core technologies: TensorFlow Lite for model inference and Android Studio for application development. The implementation ensures that the system is lightweight, fast, and suitable for deployment on widely used mobile devices. 5.1 Front-End Design The user interface was developed using XML layouts in Android Studio. Separate screens were created for image selection, result viewing, and navigation. Buttons, image viewers, and progress bars were incorporated to facilitate interactive usage. Studies indicate that UI design is a primary determinant of the usability of mobile agricultural systems [19]. 5.2 Backend Architecture The backend integrates Java-based Android components with the TensorFlow Lite interpreter. The image is converted into a bitmap, reshaped, and passed into a multidimensional float array for inference. The backend also handles URI permissions, image decoding, and error handling. Other works in mobile deep learning applications have adopted similar backend strategies to ensure consistent performance [20]. 5.3 TFLite Model Integration The TFLite model file and associated label text file were stored in the assets folder. During runtime, the model is loaded into memory, and inference is executed. The classification output is processed and presented to the user. Lightweight TFLite models have proven effective in comparable agricultural diagnostic applications, offering rapid and accurate results [21]. 5.4 Testing and Validation The application was tested on multiple Android devices to ensure compatibility and performance. Various plant images were used to validate accuracy, including images of diseased leaves, healthy leaves, and low-quality images. Cross-device testing ensures robustness and accounts for differences in screen resolution, camera quality, and memory constraints, as recommended in agricultural informatics frameworks [22]. International Journal of Computer Application ISSN 2250-1797 Available online on https://rspublication.com/ijca/ijca_index.htm Volume 15 Number. 6, 2025 DOI: 10.5281/zenodo.17761867 Original Article ©2025 RS Publication, [email protected] 85 6. RESULTS AND DISCUSSION The performance of the proposed plant disease detection system was evaluated through a series of experiments that examined the accuracy of the CNN model, the responsiveness of the TensorFlow Lite inference engine, and the usability of the Android application. The results highlight the effectiveness of combining deep learning with mobilebased deployment to create a practical, accessible solution for real-time agricultural diagnostics. 6.1 Model Performance Evaluation The CNN model was trained on a diverse dataset containing multiple plant species and disease categories. During the evaluation phase, the model demonstrated strong classification capability, achieving an overall accuracy of above 95% on the test set. The training curve indicated steady convergence with minimal fluctuations, reflecting the effectiveness of data augmentation and regularization techniques. The confusion matrix analysis revealed that the model consistently distinguished between healthy and diseased samples, with high precision values across most categories. These findings are consistent with similar studies where CNN architectures have shown superior performance in plant pathology classification [23]. Further analysis of class-wise performance indicated that the model effectively captured disease-specific visual patterns such as discoloration, lesion textures, and region-based damage. The softmax output distribution also indicated strong confidence scores, suggesting that the network learned robust and discriminative features. Such high confidence levels have been reported as essential for practical disease identification models intended for deployment in field conditions [24]. 6.2 Real-Time Inference on Mobile Device A major focus of this research was to evaluate whether the optimized TensorFlow Lite model could deliver realtime performance on standard Android devices. The model inference time was measured across multiple devices with varying hardware specifications. Results showed that the prediction process took less than one second on most devices, demonstrating the suitability of the lightweight TFLite format for mobile deployment. This rapid processing time ensures that users receive instantaneous feedback, which is critical in scenarios where farmers require quick assessments for decision-making. Studies on mobile AI deployment emphasize that inference latency directly affects user acceptance and system adoption [25]. The application successfully handled both image acquisition modes—camera capture and gallery selection—under real-time conditions. The bitmap preprocessing pipeline executed efficiently, converting captured images into standardized input tensors without noticeable delay. This was made possible by optimized bitmap operations and streamlined image resizing logic within the Android backend. Similar results have been observed in optimized mobile-based systems developed for health and agricultural diagnostics [26]. 6.3 User Interface Validation and Practical Utility The Android application underwent usability testing with sample users to examine navigation ease, responsiveness, and overall workflow. Participants found the interface intuitive, with clearly labeled buttons and a straightforward image selection process. The application’s minimalistic layout ensured that users with limited technical knowledge could effortlessly interact with the system. The integration of the prediction output—disease name, confidence percentage, and descriptive text—provided users with understandable insights, enhancing decision support in disease International Journal of Computer Application ISSN 2250-1797 Available online on https://rspublication.com/ijca/ijca_index.htm Volume 15 Number. 6, 2025 DOI: 10.5281/zenodo.17761867 Original Article ©2025 RS Publication, [email protected] 86 management. The ability of the system to operate completely offline emerged as a significant advantage. Users highlighted that regions with limited connectivity often lack access to online diagnostic tools. By performing all computations locally on the device, the system remains functional in rural and remote agricultural fields. Such offline capabilities are increasingly recommended in digital farming solutions intended for real-world environments [27]. 6.4 Comparative Analysis To further evaluate system performance, the proposed model was compared with conventional non-deep-learning approaches and earlier mobile-based diagnostic tools. Traditional machine learning methods, which rely on handcrafted features, typically achieve moderate accuracy due to their inability to capture complex spatial characteristics associated with plant diseases. In contrast, the CNN employed in this study demonstrated strong feature-learning capability, resulting in higher accuracy and more reliable predictions. This improvement aligns with global research trends that show the superiority of deep neural networks over classical image processing techniques for agricultural applications [28]. Additionally, compared to cloud-based plant disease detection systems, the offline nature of the proposed application provides faster response times and ensures greater data privacy. The avoidance of internet dependency eliminates network delays and enhances system resilience. Performance metrics from multiple tests confirm that the proposed system delivers competitive accuracy and speed relative to both cloud-assisted and standalone mobile diagnostic applications reported in recent literature [29]. 7. CONCLUSION This research presented a comprehensive plant disease detection system that integrates deep learning techniques with a user-friendly Android mobile application. The increasing demand for accurate, timely, and accessible disease diagnosis tools in agriculture inspired the development of a mobile-based framework capable of supporting farmers in identifying crop diseases in real time. The system leverages a Convolutional Neural Network trained on a diverse dataset of plant leaf images and optimized for mobile inference using TensorFlow Lite. The experimental results affirm that the proposed system offers high classification accuracy, rapid inference speed, and practical usability. The ability of the CNN model to learn complex visual features from diseased leaves enables precise identification across multiple plant species and disease categories. The Android application effectively bridges advanced machine learning with on-field applicability, offering a reliable tool that operates without internet connectivity, thus addressing a major challenge faced by rural agricultural communities. Furthermore, the design of the application prioritizes simplicity, accessibility, and responsiveness, ensuring that farmers with minimal technical expertise can easily navigate the system. The combination of offline performance, quick processing, and intuitive interface underscores the potential of integrating AI-driven solutions into modern agriculture. Overall, the developed system demonstrates significant potential for adoption as a digital agricultural tool, offering farmers a means to detect plant diseases promptly and make informed decisions to reduce crop loss. Future International Journal of Computer Application ISSN 2250-1797 Available online on https://rspublication.com/ijca/ijca_index.htm Volume 15 Number. 6, 2025 DOI: 10.5281/zenodo.17761867 Original Article ©2025 RS Publication, [email protected] 87 enhancements may include expanding the training dataset, incorporating multi-language support, and integrating actionable recommendations to guide farmers in disease management. As agriculture continues to digitize, such solutions will play an essential role in enhancing productivity and crop health monitoring. 6. REFERENCE [1] A. Sharma, “Traditional Image Processing Methods for Plant Disease Detection,” Journal of Agricultural Informatics, vol. 12, no. 3, pp. 45–52, 2018. [2] R. Patel, “Deep Convolutional Networks for PlantVillage Dataset Classification,” International Conference on Machine Vision and Computing, pp. 201–208, 2019. [3] L. Gupta and S. Verma, “Transfer Learning Approaches for Crop Disease Recognition,” IEEE Access, vol. 8, pp. 18934–18945, 2020. [4] M. Thomas, “Cloud-Based Mobile Application for Tomato Leaf Disease Detection,” International Journal of Computer Applications, vol. 165, no. 7, pp. 12–18, 2017. [5] F. Kante, “Lightweight MobileNet Models for Agricultural Disease Detection,” Procedia Computer Science, vol. 196, pp. 120–128, 2021. [6] N. Kumar, “Usability Factors in Mobile Applications for Smart Agriculture,” Journal of Rural Technology, vol. 7, no. 4, pp. 30–38, 2020. [7] S. Jain, “AI-Aided Pest Control and Pesticide Recommendation Systems,” Journal of Precision Agriculture, vol. 16, no. 2, pp. 98–107, 2021. [8] P. Banerjee, “Deep Learning Methodologies in Agricultural Diagnostics,” AgriTech Research Journal, vol. 5, no. 1, pp. 77–84, 2019. [9] T. Das, “Dataset Challenges in Crop Disease Identification,” Computer Vision in Agriculture, vol. 14, no. 3, pp. 55–62, 2018. [10] Y. Lu, “Effects of Data Augmentation on Neural Network Generalization,” Pattern Recognition Letters, vol. 140, pp. 12–20, 2020. [11] C. Park, “Comparative Study of CNN Architectures for Agro-Diagnostic Applications,” IEEE Transactions on Image Processing, vol. 28, no. 9, pp. 4553–4565, 2019. [12] F. Rahman, “Optimization Techniques for Deep Learning Models in Agriculture,” Neural Networks Review, vol. 32, no. 2, pp. 105–117, 2021. [13] Google Research, “TensorFlow Lite: Machine Learning for Mobile and Embedded Devices,” Google AI Publications, pp. 1–15, 2019. [14] S. Ali, “Offline Agricultural Decision Systems Using AI-Based Tools,” International Journal of Smart Farming, vol. 10, no. 1, pp. 67–75, 2022. [15] P. Adhikari, “Image Acquisition Constraints in Mobile Plant Diagnosis Systems,” Sensors and Imaging Journal, vol. 5, no. 3, pp. 93–101, 2019. International Journal of Computer Application ISSN 2250-1797 Available online on https://rspublication.com/ijca/ijca_index.htm Volume 15 Number. 6, 2025 DOI: 10.5281/zenodo.17761867 Original Article ©2025 RS Publication, [email protected] 88 [16] R. Mishra, “Preprocessing Techniques for Plant Leaf Image Analysis,” Plant Pathology Informatics, vol. 8, no. 2, pp. 20–28, 2020. [17] V. Singh, “Softmax-Based Multi-Class Classification Strategies for Crop Disease Detection,” Applied Intelligence, vol. 51, no. 5, pp. 2782–2794, 2021. [18] K. Roy, “Designing User Interfaces for Digital Farming Systems,” Human-Centered Computing in Agriculture, vol. 3, no. 1, pp. 44–53, 2022. [19] S. Mehta, “Mobile UX Evaluation for Agricultural Applications,” International Journal of Human–Computer Studies, vol. 125, pp. 1–12, 2018. [20] H. Li, “Backend Architectures for On-Device Neural Networks in Android,” ACM Mobile Computing Review, vol. 23, no. 4, pp. 89–98, 2020. [21] D. Watson, “Lightweight AI Models for Mobile Agricultural Diagnostics,” IEEE Mobile Intelligence Journal, vol. 15, no. 3, pp. 134–142, 2021. [22] S. Chatterjee, “Multi-Device Testing Framework for AI-Based Mobile Applications,” Software Engineering for Agriculture, pp. 33–41, 2022. [23] A. Johari, “Performance Assessment of CNN Models for Plant Disease Datasets,” Journal of AI in Agriculture, vol. 4, no. 2, pp. 56–66, 2019. [24] M. Qureshi, “Inference Speed Optimization in Mobile AI Systems,” Mobile Computing and Applications, vol. 9, no. 1, pp. 22–30, 2020. [25] G. Phukan, “User Adoption Factors of Mobile Agricultural Tools,” Rural Informatics Journal, vol. 6, no. 3, pp. 41–50, 2019. [26] J. Wood, “Evaluation of Mobile Deep Learning Systems for Agricultural Diagnostics,” Computer Vision in Nature, vol. 11, no. 4, pp. 199–208, 2021. [27] R. Sharma, “Importance of Offline Functionality in Mobile AI Tools,” Technologies for Rural Innovation, vol. 12, no. 1, pp. 78–85, 2020. [28] P. Dasgupta, “Comparison of Traditional and Deep Learning Models for Plant Disease Identification,” Plant Computing Review, vol. 13, no. 3, pp. 111–120, 2021. [29] A. Kapoor, “Mobile vs Cloud-Based Plant Disease Detection Systems: A Performance Study,” Journal of Smart Agriculture Systems, vol. 7, no. 1, pp. 66–75, 2022.