Publication: Asma yaprağında hastalık tespiti için derin öğrenme tabanlı mobil uygulama geliştirilmesi
Abstract
Asma yaprağında hastalık tespiti için derin öğrenme tabanlı mobil uygulama geliştirilmesi Teknolojinin hızla gelişmesi paralelinde bilgiye erişimi de hızlandırmıştır. Mobil (taşınabilir) cihazlar hem iletişim yönünden hem de bilgiye erişme noktasında en çok tercih edilen teknolojik cihazlardır. Bunun sonucu olarak, istenilen veriye ulaşmak geçmişe göre çok daha pratik bir durum kazanmıştır. Yaygın olarak kullanılan bu cihazlar günümüzde akıllı sistemlerin gelişmesinde de önemli rol oynamaktadır. Akıllı sistemler tarım endüstrisinde de yaygın bir biçimde kullanılmaya başlanmıştır. Gerek küresel ısınma gerekse iklim değişikliği etkisiyle tarımsal ürünlerde verim olumsuz olarak etkilenmektedir. Su seviyelerindeki azalmayla birlikte tarım alanlarındaki sulama günümüzde olduğu gibi gelecekte de kritik önemini korumaya devam edecektir. Bu çerçevede tarıma yönelik atılacak her adım gelecek için büyük bir değer taşımaktadır. Tarımla elde edilecek ürünlerin ürün kalitesi ve verimliliği oldukça önemlidir. Tarım ürünlerindeki bilinen hastalıkların yanı sıra değişen çevre faktörleriyle yeni çeşitli negatif etkenler de ortaya çıkmaktadır. Bu tip tespitlerin yapılmasını sağlamak ve gerektiğinde yerinde tespit yöntemlerinin geliştirilmesini sağlamak amacıyla yeni bir model önerisi ile farklı bir bakış açısı kazandırmak bu tez kapsamında hedeflenmektedir.Bu çalışmada sorunlar çerçevesinde son kullanıcının aktif biçimde rol alacağı, gelişmeye ve dinamik biçimde öğrenmeye açık interaktif bir sistem geliştirilmesi hedeflenmektedir. Çalışmada sunulan sistem ile birlikte asma yaprakları üzerinde hastalık tespiti yapan bir model geliştirmesi yapılmıştır. Eğitim ve test aşamalarında kullanılması için asma yaprağının görsellerinden oluşan bir veri seti elde edilmiştir. Veri seti hazırlanırken, sistemin bir mobil cihaza kurulması amaçlandığından, görseller mobil cihazlar aracılığıyla sağlanmıştır. 3122 asma yaprağı görsel verileri iPhone İşletim Sistemi (iOS) ve Android işletim sistemlerine sahip mobil cihazlarla toplanmış, böylelikle özgün bir veri seti oluşturulmuştur. Bu yapraklardan 465 tanesi model eğitim ve doğrulaması ile birlikte test verileri olarak model geliştirme aşamasında kullanılmıştır. Ön işleme süreçlerinde ve eğitim aşamasında modelden daha verimli sonuç alınabilmesi için, yaprakların bulunduğu ortamdaki karmaşıklıkla birlikte düşük kaliteye sahip ön işlem süreci başarısız olan yapraklar model eğitim ve doğrulama adımlarında kullanılmamıştır. Gerekli önişlem ve etiketleme işlemleriyle dengeli ikili sınıfsal dağılıma sahip bu veriler ileri görüntü işlemeye hazır hale getirilmiştir. Görüntü işleme teknikleri kullanılarak Python üzerinde bir makine öğrenmesi modeli geliştirilmiştir. Aynı zamanda mobil cihazlar için de bir mobil uygulama ile sistem desteklenmiştir. Mobil cihazlarda kamera üzerinden elde edilecek görselleri, sistemin asma yapraklarının sağlıklı ya da hastalık tespitini yaparak sınıflandırması hedeflenmiştir. Python vasıtasıyla oluşturulan derin öğrenme tabanlı makine öğrenmesi modeli, mobil uygulamaya dâhil edilerek aktif kullanılabilecek bir uygulama geliştirilmiştir. Python ve mobil cihazda kullanılan uygulama ile çıktı alınan model üzerinde yapılan tahminlemelerde yakın sonuçlar çıkması sebebiyle her iki ortamda da kullanılabilecek hale getirilmiştir. Çalışmalarımızda doğruluğu artırmak, farklı model yapılarını kullanarak kıyaslama olanağını artırabilecek bir sistem tasarlanmıştır. Bu kapsamda Python tabanlı yaptığımız modele ek olarak, Microsoft tarafından geliştirilen Makine Öğrenimi .NET (ML.NET) tabanlı makine öğrenmesi modeliyle de tasarladığımız modeli güçlendirmek hedeflenmiştir. ASP.Net Core web uygulama çerçevesi kullanılarak bir web uygulama hazırlanmıştır. ML.NET modelinde 3122 fotoğrafın tamamı ön işleme tabi tutulmadan orjinal veri yapısı ile eğitime tabi tutulmuştur. Mobil uygulama tarafında aynı resim üzerinden Python modeliyle birlikte ML.NET üzerinden de eş zamanlı sınıflandırma yapılabilmektedir. Uygulamalar arasında haberleşmenin sağlanabilmesi için Python ve ASP.Net Core tabanlı gerekli hizmet bağlantılarıyla birlikte mobil uygulamanın çevrimiçi ortamlarla iletişimi sağlanmıştır.Üzüm yapraklarındaki bu hastalık zamanında tespit edilemezse ileriye yönelik üzümlerin küflenmesine, asma yapraklarının ticari kullanılamamasına ve asma yaprağıyla birlikte üzümün zarar görmesine yol açacağından ürünlerin tüketilemez duruma gelmesine sebep olacaktır. Uygulama tarafında tahmin sonuçlarının kullanıcıya gösterilerek, kullanıcıdan geri dönüş alınması sağlanmıştır. Bu kapsamda son kullanıcının tarım ürünlerindeki kalite ve verimi dolaylı olarak tespit etmesi amaçlanmıştır. Bu amaçla birlikte elde edilen verilerle çalışan modelin tekrar tekrar kendini güncelleyebilir olmasına da katkı sağlaması hedeflenmiştir.Bu çalışmadaki hazırlanan sistemin, farklı tarım ürünlerinde verdiği sonuçları da görebilmek için farklı çalışmalarda kullanılmış bir hazır veri seti üzerinde biber yapraklarıyla uygulamada kıyaslama yapılmıştır. Çalışmada yapmış olduğumuz uygulamalara benzer oranlarda eğitim, doğrulama ve test verileri verilerek sonuçlar gözlemlenmiştir. Arka planı temizlenmiş sadece yaprak görüntülerinin olduğu 2475 biber yaprağı görseli uygulamada başarılı bir şekilde eğitimi tamamlamıştır. Hazır görseller üzerinde sınıflandırma işlemini başarıyla uygulayabilmektedir. Bu analiz ile sadece ürün değişimi yapılarak uygulamanın farklı veri setleri ile çalışabilirliği deneyimlenmiştir. Bu kapsamda, tarım alanında yapılacak ilerideki çalışmalara yönelik geliştirilmesi hedeflenen modellere, diğer tarım ürünlerine de uygulanabilirliği ölçüsünde yön vermesi ve bu yönde çalışmalara katkı sağlaması beklenmektedir. Bu çalışma ile birlikte mobil uygulamanın makine öğrenmesine yönelik yapılacak çalışmalarda daha öncü rol oynayacağı düşünülmektedir.
Development of a deep learning based mobile application for disease detection in grapevine leaves The rapid development of technology has accelerated access to information in parallel. Mobile (portable) devices are the most preferred technological devices both in terms of communication and access to information. As a result, accessing the desired data has become much more practical than in the past. These widely used devices also play an important role in the development of smart systems today. Intelligent systems have started to be widely used in the agricultural industry. Both global warming and climate change have negatively affected the yield of agricultural products. With the decrease in water levels, irrigation in agricultural areas will continue to maintain its critical importance in the future as it is today. In this context, every step to be taken towards agriculture is of great value for the future. Product quality and productivity of the products to be obtained through agriculture are very important. In addition to the known diseases in agricultural products, various new negative factors are emerging with changing environmental factors. This thesis aims to provide a different perspective with a new model proposal in order to ensure that such determinations are made and to develop on-site detection methods when necessary. In this study, it is aimed to develop an interactive system in which the end user will take an active role within the framework of the problems, open to development and dynamic learning. With the system presented in the study, a model for disease detection on grapevine leaves was developed. A dataset consisting of images of grapevine leaves was obtained for use in the training and testing phases. While preparing the dataset, the images were provided via mobile devices since the system was intended to be installed on a mobile device. 3122 grapevine leaf visual data were collected using mobile devices running the iPhone Operating System (iOS) and Android operating systems, thereby creating a unique dataset. 465 of these leaves were used in the model development phase as test data along with model training and validation. In order to obtain more efficient results from the model during preprocessing and training, leaves that failed the preprocessing step due to low quality and complexity in their environment were not used in the model training and validation steps. With the necessary preprocessing and labeling, these data with a balanced binary class distribution were made ready for further image processing. A machine learning model was developed in Python using image processing techniques. At the same time, the system was supported with a mobile application for mobile devices. It is aimed for the system to classify the images to be obtained through the camera on mobile devices by detecting healthy or diseased vine leaves. A deep learning-based machine learning model created through Python was included in the mobile application and an application that can be used actively was developed. Since the predictions made on the model outputted with Python and the application used on the mobile device yielded close results, it has been made usable in both environments. We designed a system that can increase the accuracy of our studies and increase the possibility of comparison by using different model structures. In this context, in addition to our Python-based model, we aimed to strengthen the model we designed with the Machine Learning .NET (ML.NET) based machine learning model developed by Microsoft. A web application was prepared using the ASP.Net Core web application framework. In the ML.NET model, all 3122 photos were trained with the original data structure without preprocessing. On the mobile application side, simultaneous classification can be performed on the same image via ML.NET as well as Python model. In order to provide communication between the applications, Python and ASP.Net Core based service connections were used to enable the mobile application to communicate with online environments. If this disease in grape leaves is not detected in time, it will cause the grapes to mold, the grape leaves cannot be used commercially and the grape will be damaged along with the grape leaves, which will cause the products to become inedible. On the application side, the prediction results are shown to the user and feedback from the user is provided. In this context, it is aimed for the end user to indirectly determine the quality and yield of agricultural products. For this purpose, it is also aimed to contribute to the fact that the model working with the data obtained together can update itself repeatedly.In order to see the results of the system prepared in this study on different agricultural products, a comparison was made in the application with pepper leaves on a ready-made dataset used in different studies. The results were observed by giving training, validation and test data at similar rates to the applications we have made in the study. In the application, 2475 pepper leaf images with only leaf images with cleaned background completed the training successfully. It can successfully apply the classification process on ready-made images. With this analysis, only the product change was made and the application's ability to work with different datasets was experienced. In this context, it is expected that the models to be developed for future work in the field of agriculture will guide other agricultural products to the extent that they are applicable and contribute to work in this direction. With this study, it is thought that the mobile application will play a more pioneering role in future studies on machine learning.
Development of a deep learning based mobile application for disease detection in grapevine leaves The rapid development of technology has accelerated access to information in parallel. Mobile (portable) devices are the most preferred technological devices both in terms of communication and access to information. As a result, accessing the desired data has become much more practical than in the past. These widely used devices also play an important role in the development of smart systems today. Intelligent systems have started to be widely used in the agricultural industry. Both global warming and climate change have negatively affected the yield of agricultural products. With the decrease in water levels, irrigation in agricultural areas will continue to maintain its critical importance in the future as it is today. In this context, every step to be taken towards agriculture is of great value for the future. Product quality and productivity of the products to be obtained through agriculture are very important. In addition to the known diseases in agricultural products, various new negative factors are emerging with changing environmental factors. This thesis aims to provide a different perspective with a new model proposal in order to ensure that such determinations are made and to develop on-site detection methods when necessary. In this study, it is aimed to develop an interactive system in which the end user will take an active role within the framework of the problems, open to development and dynamic learning. With the system presented in the study, a model for disease detection on grapevine leaves was developed. A dataset consisting of images of grapevine leaves was obtained for use in the training and testing phases. While preparing the dataset, the images were provided via mobile devices since the system was intended to be installed on a mobile device. 3122 grapevine leaf visual data were collected using mobile devices running the iPhone Operating System (iOS) and Android operating systems, thereby creating a unique dataset. 465 of these leaves were used in the model development phase as test data along with model training and validation. In order to obtain more efficient results from the model during preprocessing and training, leaves that failed the preprocessing step due to low quality and complexity in their environment were not used in the model training and validation steps. With the necessary preprocessing and labeling, these data with a balanced binary class distribution were made ready for further image processing. A machine learning model was developed in Python using image processing techniques. At the same time, the system was supported with a mobile application for mobile devices. It is aimed for the system to classify the images to be obtained through the camera on mobile devices by detecting healthy or diseased vine leaves. A deep learning-based machine learning model created through Python was included in the mobile application and an application that can be used actively was developed. Since the predictions made on the model outputted with Python and the application used on the mobile device yielded close results, it has been made usable in both environments. We designed a system that can increase the accuracy of our studies and increase the possibility of comparison by using different model structures. In this context, in addition to our Python-based model, we aimed to strengthen the model we designed with the Machine Learning .NET (ML.NET) based machine learning model developed by Microsoft. A web application was prepared using the ASP.Net Core web application framework. In the ML.NET model, all 3122 photos were trained with the original data structure without preprocessing. On the mobile application side, simultaneous classification can be performed on the same image via ML.NET as well as Python model. In order to provide communication between the applications, Python and ASP.Net Core based service connections were used to enable the mobile application to communicate with online environments. If this disease in grape leaves is not detected in time, it will cause the grapes to mold, the grape leaves cannot be used commercially and the grape will be damaged along with the grape leaves, which will cause the products to become inedible. On the application side, the prediction results are shown to the user and feedback from the user is provided. In this context, it is aimed for the end user to indirectly determine the quality and yield of agricultural products. For this purpose, it is also aimed to contribute to the fact that the model working with the data obtained together can update itself repeatedly.In order to see the results of the system prepared in this study on different agricultural products, a comparison was made in the application with pepper leaves on a ready-made dataset used in different studies. The results were observed by giving training, validation and test data at similar rates to the applications we have made in the study. In the application, 2475 pepper leaf images with only leaf images with cleaned background completed the training successfully. It can successfully apply the classification process on ready-made images. With this analysis, only the product change was made and the application's ability to work with different datasets was experienced. In this context, it is expected that the models to be developed for future work in the field of agriculture will guide other agricultural products to the extent that they are applicable and contribute to work in this direction. With this study, it is thought that the mobile application will play a more pioneering role in future studies on machine learning.