Lightweight Deep Learning Models for Leaf-Based Irrigation Level Classification
Year: 2026
Type: preprint
Abstract: Effective irrigation management plays a vital role in precision farming, as it has a direct impact on both crop productivity and water usage efficiency. This study introduces a lightweight deep learning–based method for identifying irrigation levels from leaf images. Rather than relying on deep and computationally intensive models commonly used in earlier research, this work examines whether compact, task-focused convolutional neural networks can deliver equal or better performance. Three lightweight architectures are developed and assessed: a Simple Convolutional Neural Network, a Depthwise Separable CNN, and a Mini-Inception model inspired by multi-scale feature extraction. The proposed models undergo training using a leaf image dataset covering four irrigation conditions (25%, 50%, 75%, and 100%). Each network is carefully designed to reduce the number of trainable parameters while still learning meaningful and distinguishing features. Evaluation outcomes reveal that all three models achieve classification accuracies above 99%, despite a substantially reduced parameter count relative to traditional deep learning techniques. Notably, the Mini-Inception model reaches an accuracy of 99.86% with roughly 0.2 million parameters, outperforming larger benchmark models reported in earlier studies. These results suggest that irrigation-induced stress characteristics in leaf images can be reliably captured using lightweight network designs, leading to lower computational cost, reduced overfitting, and improved generalization. This leads to the proposed approach being highly suitable for real-time and edge-based agricultural applications, including mobile platforms and IoT-enabled smart irrigation systems.
Source: Research Square
Institution: Visvesvaraya Technological University
Language: English
Topic: Smart Agriculture and AI
Subfield: Plant Science
Domain: Life Sciences
Open Access status: green
OpenAlex ID: W7235336224
Cites:
Cited by:
Related to:
