[Paper Review] Artificial Intelligence based drone for early disease detection and precision pesticide management in cashew farming

Manoj Kumar Rajagopal, Bala Murugan MS|arXiv (Cornell University)|Mar 15, 2023

Smart Agriculture and AI8 citations

TL;DR

The paper presents an edge-AI UAV system that detects cashew leaf disease (anthracnose) and guides precision pesticide application, achieving high classification accuracy using MobileNetV2 and PlantVillage data.

ABSTRACT

The use of unmanned aerial vehicles (UAV) is revolutionizing the agricultural industry. Cashews are grown by approximately 70% of small and marginal farmers, and the cashew industry plays a critical role in their economic development. To take timely counter measures against plant diseases and infections, it is imperative to monitor and detect diseases as early as possible and take suitable measures. Using UAVs, such as those that are equipped with artificial intelligence, can assist farmers by providing early detection of crop diseases and precision pesticide application. An edge computing paradigm of Artificial Intelligence is employed to process this image in order to make decisions with the least amount of latency possible. As a result of these decisions, the stage of infestation, the crops affected, the method of prevention of spreading the disease, and what type and amount of pesticides need to be applied can be determined. UAVs equipped with sensors detect disease patterns quickly and accurately over large areas. Combined with AI algorithms, these machines can analyse data from a variety of sources such as temperature, humidity, CO2 levels and soil composition. This allows them to recognize disease symptoms before they become visible. Early detection allows for more effective control strategies that can reduce costs caused by lost production due to infestations or crop failure. Using an end-to-end training architecture, mobileNetV2 determines how to classify anthracnose disease in cashew leaves. A standard PlantVillage dataset is used for performance evaluation and for standardization. Additionally, samples captured with a drone present a variety of image samples captured in a variety of conditions, which complicates the analysis. According to our analysis, we were able to identify the anthracnose with 95% accuracy and the healthy leaves with 99% accuracy.

Motivation & Objective

• Motivate early detection of cashew diseases to prevent spread and crop loss.
• Develop an edge computing AI pipeline on a UAV to minimize latency in decision-making.
• Enable targeted pesticide application based on disease detection and environmental sensor data.
• Utilize standardized datasets to evaluate performance and extend to real-world drone-captured samples.

Proposed method

• End-to-end training architecture using MobileNetV2 for classifying anthracnose on cashew leaves.
• Use of a PlantVillage dataset for performance evaluation and standardization.
• Drone-mounted sensors capture multisource data to aid disease pattern recognition (temperature, humidity, CO2, soil).
• Edge computing paradigm to minimize latency in disease detection decisions.
• Ability to determine infestation stage, affected crops, prevention methods, and pesticide type/amount.

Experimental results

Research questions

• RQ1 Can an edge-AI UAV accurately detect anthracnose on cashew leaves in real-world drone imagery?
• RQ2 What is the classification accuracy for anthracnose and healthy leaves using the proposed end-to-end MobileNetV2 approach?
• RQ3 How well does the system generalize from PlantVillage data to drone-captured cashew leaf samples under varying conditions?
• RQ4 What pesticide management guidance can be derived from detected disease state and environmental sensor data?

Key findings

• Anthracnose on cashew leaves was identified with 95% accuracy.
• Healthy leaves were identified with 99% accuracy.
• An edge computing workflow enables low-latency processing for on-board decision making.
• The system leverages a standard PlantVillage dataset for benchmarking and validation.
• Drone-captured samples introduce variability; the approach addresses analysis under diverse conditions.