Implementing a Novel Multimodal Neural Network Approach using Dynamic Hyperparameter Selection within an Unmanned Aerial Vehicle for the Early Detection of Crop Diseases
With global food demand rising and 10–20% of crops lost to disease during production and harvest, early detection is critical to reducing agricultural losses and improving food security. To mitigate this issue, I created an autonomous crop disease detection system that combines an Unmanned Aerial Vehicle (UAV) with a multimodal neural network capable of analyzing both aerial data and close-up images and videos of crops. This project was developed as an extension of Ceres (see below). The model was trained using a dataset of over 2,000 expert-validated inputs spanning 15 diseases, with the data divided into 80% for training and 20% for testing. The image-processing branch of the network consists of nine layers across six distinct layer types, while the video-processing branch incorporates six layers. I also developed a dynamic hyperparameter selection system that automatically adjusts model parameters during training based on real-time accuracy. The network was then integrated directly into the UAV system, creating an autonomous device capable of surveying fields and identifying potential disease outbreaks. By combining artificial intelligence, computer vision, and aerial robotics, this project demonstrates how advanced technology can enable scalable, efficient, and early disease detection to support global food security.
Ceres: A Novel Device Utilizing Raspberry Pi & Neural Networks to Detect Crop Diseases Using Imaging
Ceres is an artificial intelligence-powered crop disease detection system designed to make disease diagnosis faster, more accessible, and more affordable for growers. As global food demand continues to rise, crop diseases cause significant agricultural losses, with 10–20% of crops lost during production and harvest, making early detection essential to improving food security. To address this challenge, I developed a neural network-based model that can identify crop diseases directly in the field, eliminating the need for time-consuming laboratory testing and expert validation. The model was trained on approximately 13,000 expert-validated images spanning 14 crop types and 24 diseases, utilizing a 12-layer architecture with image preprocessing techniques such as median filtering and feature extraction. When tested on previously unseen images, the system achieved an accuracy of 89%. To make the technology practical for real-world use, I integrated the model into a 3D-printed portable device consisting of a Raspberry Pi, camera, and LCD. Users simply capture an image of a plant, and the device analyzes it and returns a diagnosis within seconds. By combining machine learning with accessible hardware, Ceres demonstrates how technology can help improve crop management, reduce food waste, and support global food security efforts!
Implementing a Partial Least Squares Regression Model for the Precise Prediction of Soil Nutrients Based on Hyperspectral Reflectance Data
Soil nutrient deficiencies are becoming a growing problem in agriculture with increasing global food demand. Predicting nutrient percentages can help farmers better manage soil health, improve crop yields, and make more informed agricultural decisions. In this project, developed in collaboration with The Ohio State University AgSensing Lab, I designed a predictive modeling approach that uses a combination of Partial Least Squares Regression (PLS) and Ridge Regression to predict soil nutrient percentages based on spectral reflectance data. The dataset was split into 80% training data and 20% testing data. I preprocessed the data using feature scaling, and cross-validation was applied to optimize the number of latent components in the model. The model was then trained using the processed training dataset. During training, the model achieved an average R-squared value of 0.94 and a Root Mean Square Error (RMSE) of 0.04. After training, the model was evaluated on the testing dataset, where it was given data values it had not seen before in order to verify its accuracy and generalization ability. On the test data, the model achieved a final R-squared value of 0.87 and a Root Mean Square Error of 0.05. These results indicate that the model performs effectively when applied to new, unseen spectral reflectance data. Overall, this project demonstrates that a Partial Least Squares Regression model combined with Ridge Regression can be developed for the accurate prediction of nutrient deficiencies, with potential applications in improving crop yields and helping meet increasing global food demand.