Problem/Objective

An agro-chemical company was seeking a partner to understand the feasibility of measuring the efficacy of agri-inputs on corn plants by estimating the size of the leaves of the corn plant. Corn develops in a progressive manner: the area of the leaf of the plant is used to determine the health of the plant and unhealthy plants appear to have narrow and twisted leaves compared to healthy plants. The underlying goal was to measure the leaf area of corn plants. However, the manual process was extremely time consuming and made the process of measuring leaves across large swathes of plants untenable.

Solution/Approach

The agrochemical company collaborated with Tiger Analytics to utilize drone images of the farm with deep-learning-based image processing techniques to estimate the leaf area index. The goal was to utilize drone imagery of the farm and determine growth parameters of individual plants and compare the produce of the farm regions that had different levels of fertilizer utilization. Tiger Analytics drove the project where drones were equipped with ZENMUSE cameras to capture high-quality images of the farm. They also added markers on the ground to identify crop sections pertaining to those rows of corn plants that had to be examined. To have sufficient samples to train the deep learning models, Tiger Analytics stimulated training data using an open-source 3D computer graphics software - Blender. The finely tuned models then enabled the study of discriminative features that operated reliably on real-world images. PyTorch was used to build deep learning models.

Impact/Implementation

During the implementation process, plots of land were monitored every week. To overcome any problem related to the use of deep learning models, data was synthetically created with deep knowledge of vegetation science so as to simulate real-world data, and accurately capture leaf area index based on the simulation parameters.

The agrochemical company was able to test the efficacy of different products across a large sample on the farms. The farm was divided into regions with different treatments - the growth parameters were accurately measured across several weeks and were fed back to the product research and development teams for analysis. The solution consisted of a unique and detailed method to create synthetic images, a robust and generalizable deep learning inference, strategic image capture and scoring pipeline with a clear path to derive farming intelligence from sensors. Outcome in terms of mean absolute error in identifying the images of diseased or affected crops was drastically lower for Synthetic test images.