[Paper Review] Farmer's Assistant: A Machine Learning Based Application for Agricultural Solutions

Shloka Gupta, Akshay Chopade+2 more|arXiv (Cornell University)|Apr 24, 2022

Smart Agriculture and AI4 citations

TL;DR

This paper presents Farmer's Assistant, a web-based machine learning application that provides crop and fertilizer recommendations using Random Forest and rule-based models, respectively, and detects plant diseases via an EfficientNet model achieving 99.8% validation accuracy. The system enhances interpretability using LIME to highlight disease-affected leaf regions, offering actionable insights for farmers in resource-constrained settings.

ABSTRACT

Farmers face several challenges when growing crops like uncertain irrigation, poor soil quality, etc. Especially in India, a major fraction of farmers do not have the knowledge to select appropriate crops and fertilizers. Moreover, crop failure due to disease causes a significant loss to the farmers, as well as the consumers. While there have been recent developments in the automated detection of these diseases using Machine Learning techniques, the utilization of Deep Learning has not been fully explored. Additionally, such models are not easy to use because of the high-quality data used in their training, lack of computational power, and poor generalizability of the models. To this end, we create an open-source easy-to-use web application to address some of these issues which may help improve crop production. In particular, we support crop recommendation, fertilizer recommendation, plant disease prediction, and an interactive news-feed. In addition, we also use interpretability techniques in an attempt to explain the prediction made by our disease detection model.

Motivation & Objective

• To address the lack of accessible, accurate, and interpretable agricultural decision support tools for smallholder farmers in India.
• To develop an open-source, user-friendly web application integrating crop recommendation, fertilizer selection, and plant disease detection using machine learning.
• To improve model interpretability in plant disease detection using LIME, enabling farmers and researchers to understand prediction logic.
• To reduce crop loss due to disease and suboptimal farming practices through automated, data-driven recommendations.
• To enhance model generalization by identifying data gaps and planning for future data expansion and improved segmentation.

Proposed method

• Utilizes the EfficientNet deep learning model fine-tuned on a leaf image dataset for plant disease classification, achieving 99.8% validation accuracy.
• Employs a Random Forest model trained on soil (N, P, K, pH) and weather features to recommend suitable crops.
• Applies a rule-based classification system to recommend fertilizers based on NPK values and crop requirements.
• Integrates web scraping to deliver real-time, relevant agricultural news to users via a dynamic news feed.
• Employs the LIME interpretability technique to generate local explanations by highlighting salient regions in leaf images that influence disease predictions.
• Designs a modular web application interface to allow farmers to input soil data, upload leaf images, and receive instant recommendations and explanations.

Experimental results

Research questions

• RQ1 Can a unified, open-source web application effectively integrate crop recommendation, fertilizer selection, and plant disease detection using machine learning?
• RQ2 How accurately can a pre-trained EfficientNet model detect plant diseases from leaf images in a real-world agricultural context?
• RQ3 To what extent can LIME provide reliable and meaningful explanations of model predictions in plant disease detection?
• RQ4 How can interpretability techniques improve trust and usability of ML models for non-technical farmers?
• RQ5 What are the key limitations in model generalization, and how can future data collection and annotation strategies address them?

Key findings

• The EfficientNet model achieved a validation accuracy of 99.8% on the plant disease detection task, demonstrating high performance on the chosen dataset.
• LIME explanations successfully highlighted diseased regions in leaf images, confirming that the model focuses on biologically relevant areas.
• The rule-based fertilizer recommendation system provides consistent, logic-driven suggestions based on NPK values and crop needs.
• The Random Forest model for crop recommendation effectively leverages soil and weather data to generate context-aware suggestions.
• The application demonstrates strong usability and performance in a real-world deployment scenario, with smooth inference and responsive interface.
• Limitations include poor generalization to out-of-domain images and the need for expanded, diverse training data to improve robustness.