Agricultural Vehicle Image Sampling for ML Training
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current agricultural spray systems face challenges in accurately detecting target plants due to limited training data and false positives/negatives, which affects the precision of spraying operations.
Innovation Solution
A system equipped with cameras and a trained machine-learning model on an agricultural vehicle that captures images, applies image-selection parameters to sample and store images for machine-learning training, including those with and without target plants, to improve detection accuracy and troubleshoot the system.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If all captured images are stored for machine-learning training, then the training data quantity increases, but the storage requirements and data processing complexity increase
Solution Approach 1:
The system extracts and stores only the most valuable images for training purposes. Image quality metrics (sharpness, exposure, contrast) and detection confidence scores are used to identify and extract only those images that will most effectively improve model performance, discarding redundant low-quality or clearly classified images.
Solution Approach 2:
Different selection criteria are applied to different subsets of images based on their characteristics. High-confidence easy samples are selected using one criterion, while low-confidence or borderline cases are selected using different criteria. Image quality thresholds are adjusted locally based on scene complexity and detection difficulty.
2Measurement precision
If image quality thresholds are set high to ensure training data quality, then detection accuracy improves, but the number of usable training images decreases
Solution Approach 1:
The system dynamically adjusts image quality thresholds and selection parameters based on the current dataset characteristics, model performance metrics, and scene conditions. Thresholds are not fixed but adapt as the model improves and as different field conditions are encountered, optimizing the balance between quality and quantity.
Solution Approach 2:
The system initially applies stringent quality filters to ensure high-quality training data, then gradually relaxes thresholds as the model accumulates more training examples. This allows the system to start with a smaller high-quality dataset and expand to include more diverse examples as confidence in the selection process increases.
3Measurement precision
If the machine-learning model is continuously retrained with new images, then detection accuracy improves, but the computational time and processing resources increase
Solution Approach 1:
Instead of continuous retraining, the system implements periodic model updates at predetermined intervals or after accumulating a threshold number of new training images. This allows batch processing of training data rather than continuous incremental learning, reducing computational overhead while maintaining model freshness.
Solution Approach 2:
New images are pre-processed, annotated, and validated in advance before being added to the training dataset. Image quality assessment and relevance filtering are performed beforehand, so that when retraining occurs, the model receives pre-prepared high-quality data, reducing the actual training time required.
4Reliability
If strict image-selection parameters are applied to filter images, then the quality of stored images improves, but the number of images available for training decreases
Solution Approach 1:
Image selection parameters are made dynamic rather than static. The system adjusts selection stringency based on current model performance, dataset diversity metrics, and field conditions. When the model struggles with certain plant types or conditions, the system automatically relaxes related selection criteria to gather more diverse training examples.
Data Source
AI summary
A system includes an agricultural vehicle, one or more cameras in mechanical communication with the agricultural vehicle, and a computer in electrical communication with the cameras. The computer is programmed to automatically analyze each image for a presence of at least one target plant using a trained machine-learning model, the trained machine-learning model having been trained with first images that include the at least one target plant and second images that do not include the at least one target plant; automatically detect, using the trained machine-learning model, the at least one target plant in a subset of the images; apply an image-selection parameter to the subset of the images to select one or more images for storage; and store the one or more images for machine-learning training in a computer storage device operably coupled to the one or more microprocessors.