An industrial agricultural producer in Romania faces a critical operational challenge: detecting diseases, pests and nutrient deficiencies in multiple crops at an early stage, before they significantly affect production. The traditional process involves manual visual inspection of crops by agronomists — slow, expensive and dependent on the availability of field experts. In an industrial production context, with large areas and tight deadlines for phytosanitary interventions, the margin for error was minimal.
The main technical challenge was building a system capable of analyzing images photographed under various conditions — natural light, different angles, variable camera quality — and producing an accurate diagnosis in real time, directly from the field, without connection to a human expert.
Nenos has developed a crop disease diagnostic platform, exposed as a ready-to-use API and integrated into an application available on both mobile and web. The farmer or agronomist takes a photo of the affected plant, and the system returns a complete diagnosis — disease, pest or nutrient deficiency identified, estimated severity and recommended solutions — in less than 30 seconds.
The core of the platform is a deep learning algorithm developed in Python, trained to recognize 68 types of diseases and damage on multiple crops. The model analyzes the image and identifies visual patterns specific to each disease — discoloration, lesions, abnormal textures, signs of infestation — regardless of the plant species or shooting conditions. In addition to diagnosis, the system offers operational recommendations based on image analysis: the optimal time to herbicide, treat or harvest.
The solution was deployed as an API on Amazon SageMaker, ensuring automatic scalability and high availability for simultaneous field use by multiple operators.
The implementation of the platform has transformed the way the producer manages crop health:
- Over 94% accuracy in identifying diseases, pests and nutrient deficiencies across multiple crops.
- Complete diagnosis returned in less than 30 seconds after photographing the plant.
- 68 types of diseases, pests and nutrient deficiencies covered in one system.
- Eliminating dependence on physical human expertise in the field for initial diagnosis.
- Reducing reaction time to the appearance of diseases, with direct impact on the quality and volume of production.
The platform is built in Python with a machine learning stack specialized in computer vision: TensorFlow and Keras for defining and training convolutional neural networks, OpenCV for image preprocessing — normalization, resizing, contrast correction, and removal of photography artifacts — and Scikit-learn for classification and model performance evaluation components.
The detection model is a convolutional neural network (CNN) trained on a representative dataset of affected crop images, covering 68 classes of diseases and damages across multiple species. The training included extensive data augmentation — rotations, flips, brightness variation, blur simulation — to ensure robustness in real-world field photography conditions, where image quality varies significantly. The final model accuracy exceeds 94%, obtained through an iterative process of training, validation, and retraining on extensive datasets.
The solution is exposed as a REST API deployed on Amazon SageMaker, which manages automatic scaling of inference based on request volume. The client application — available on both mobile and web — sends the image to the API and displays the diagnosis and recommendations in less than 30 seconds, usable directly from the field without technical knowledge.
