AGRICULTURAL ASSESSMENT USING AERIAL-IMAGERY DATA WITH DEEP CONVOLUTION NETWORK AND PRESCRIPTION MAP ALGORITHM FOR OPTIMIZED SPRAYING PATH

Abstract

This work presents the use of unmanned aerial vehicles (UAVs) within precision agriculture for weed segmentation and variable rate application (VRA) in precision agriculture. Conventional spraying methods often lack precision, leading to environmental and economic concerns. Leveraging deep learning techniques, we propose a solution to tackle these challenges efficiently. This research addresses these issues using drone imagery and deep-learning techniques for site-specific weed management. Our work combines Semantic segmentation of the crop fields and prescription map analysis to model itself as an end-to-end detection system for field surveillance. The Deep Convolutional Networks are employed to predict weed infestation in the field. We employed three diverse datasets, including one collected under Indian conditions, ensuring the robustness and applicability of our approach across different agricultural settings. Using semantic mapping, we calculated prescription maps for specific land patches, determining the spraying required in each area. This data is then utilized to guide a UAV sprayer along an optimized spraying path for the field. By planning an optimized trajectory, the sprayer can efficiently travel the shortest distance while delivering the necessary amount of spray in each region. This study also includes a comparative analysis of three distinct path-planning approaches (1) the lawnmower algorithm, (2) the Priority-based algorithm, and (3) the TSP-based algorithm to determine their suitability for optimizing field coverage and resource management in drone-based missions. Using the RGB-based data the model achieved an overall accuracy of 92.22% whereas with modified data it achieved 95.04%. Experimental results showed that both deep learning models led to high segmentation accuracy. Still, the model trained on the transformed data had accuracy 2.82% higher than the RGB data-based model. Our analysis focused on priority-based and Traveling Salesman Problem (TSP)-based methods, among others. The findings demonstrate that both priority-based and Traveling Salesman Problem (TSP)-based methods show promising results in optimizing field coverage and resource utilization.