<![CDATA[Vegetation density classification from UAV imagery is a practical necessity in fire-prone landscapes, since fuel load on the ground directly informs risk management decisions. Convolutional neural networks handle this classification reasonably well, but good performance requires careful hyperparameter tuning, and manual trial and error produces results that are unstable and hard to reproduce. This study proposes LDW-FOX, a modified FOX metaheuristic using a Linearly Decreasing Weight mechanism to automate hyperparameter tuning for an Inception-based CNN. The original FOX algorithm applies fixed movement weights throughout optimization, causing search to stagnate early. LDW-FOX gradually reduces exploration intensity across iterations, pushing search toward exploitation as it converges. Five hyperparameters, namely learning rate, dropout rate, hidden layer size, activation function, and optimizer, were tuned on a balanced 3,000 image UAV dataset spanning three vegetation density classes. Manual tuning peaked at 61.00 percent test accuracy but varied considerably across epoch settings. LDW-FOX reached a peak test accuracy of 82.48 percent and a mean of 58.47 percent, outperforming the original FOX, whose mean was 55.30 percent. LDW-FOX showed a more consistent training-test gap than other swarm-based methods, with LDW variants beating unmodified counterparts under equal budgets. High variance across configurations indicates broader generalization needs testing.]]>
Copyrights © 2026
