Explainable Weed Detection in Sugarcane Using Lightweight CNN and Grad-CAM

Weed invasion has a major negative effect on sugarcane production by decreasing the crop yield to be as low as 50 % and increasing the labor costs per acres to manually monitor the field. We propose a lightweight Convolutional Neural Network (CNN) that will enable binary classification of sugarcane field images as Crops or Weeds using a CPU. The CNN will also include the addition of Gradient-weighted Class Activation Mapping (Grad-CAM) to enhance understanding of which features of the images contributed to the classification decision made by the CNN by producing visual heatmaps. Users will have increased confidence in the automated classification because the heatmaps will show the discriminative parts of the image that determined the classification decision. A Tkinter-based Graphical User Interface (GUI) allows images to be uploaded in real-time, predicted (latency of$\sim \mathbf{0. 9 8}-\mathbf{1. 0 1 s}$) and Visualized using Grad-CAM on standard CPUs with no GPU acceleration. The system achieved 97.3% classification accuracy, 97.3% macro-precision, 97.3 % macro-recall, and 97.3% macro-F1 on the test dataset. All system-level tests passed demonstrating 100% functional reliability. The proposed system is a cost effective, interpretable, deployable solution that uses precision weed management in sugarcane growing operation.