UAV-assisted Xception-enhanced quantized CNN framework for forest-wild fire detection and impact assessment

Abstract

The proposed work presents a convolutional neural network, hereafter referred to as CNN4WFD, based on transfer learning and built on an improved version of an Xception architecture, which is used to detect wildfires. A full preprocessing pipeline (comprising denoising, normalization, and class-balanced augmentation) is also built into the architecture. This pipeline will help to alleviate data imbalance and complement the modeling power. An Xception model trained on ImageNet moves the generalized hierarchical feature representations and then fine-tunes them on wildfire-specific images. This strategy alleviates the absence of labeled data, accelerates convergence, and creates strong optimisation. The training is performed based on a categorical cross-entropy loss using the Adam optimiser and supplemented by a dynamically scheduled learning rate that improves the adaptation of the gradients and stabilizes the optimisation process. Accuracy, precision, recall, and F1-score, and inference latency support the reliability and computational efficiency of the model. CNN4WFD achieved the highest performance: a precision of 99%, recall of 100% and an F1-score of 99% in the Fire class; 100%, 99% and 99% in the No-Fire class. Both macro- and micro-averaged measures converged to 99 predictive power. CNN4WFD achieved strong classification performance on both wildfire datasets. For Dataset I, the proposed model achieved a validation accuracy of 99.74%. For Dataset II, the model achieved an independent held-out test accuracy of 98.37%, with 483 of 491 samples correctly classified. The component-wise ablation study demonstrated a cumulative improvement from 92.14% for the baseline CNN to 98.37% for the complete CNN4WFD configuration. Post-training quantization reduced model size by 74.7% and reduced inference latency from 36.5 ms to 14.8 ms while maintaining a comparable classification accuracy of 98.37%. The CNN4WFD framework achieved an overall classification accuracy of 98.37%, with a precision of 98.51%, recall of 98.37%, and F1-score of 98.35%. Furthermore, a Cohen’s kappa coefficient of 97.58% demonstrates near-perfect agreement between predicted and actual class labels of dataset II. Ablation analysis demonstrates a cumulative 6.23% accuracy improvement (from 92.14% to 98.37%) through the proposed CNN4WFD components. Quantization further achieves a 74.7% model size reduction and 2.47\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\times$$\end{document} faster inference while maintaining 98.37% classification accuracy. Whereas such close-to-perfect metrics emphasise the discriminative effectiveness of the Xception backbone, it can be argued to some extent that they can be explained by data-specific limitations. Due to its scalable and lightweight architecture, CNN4WFD can be deployed in real-time on unmanned aerial vehicles and IoT-enabled sensor networks, as well as edge devices, outperforming traditional CNN baselines both in accuracy and computational efficiency.