Early Detection and Segmentation of Bark Beetles Using ResNet50

Abstract

Climate change poses a significant threat to
global ecosystems, leading to an increased proliferation
of pests in both agricultural and forested regions. In
Mexico, this phenomenon has facilitated the spread
of the bark beetle (as Dendroctonus) to higher
altitudes, a challenge that traditional manual monitoring
methods cannot effectively address. Currently,
forest rangers manually count beetles in traps, a
time-consuming process that often leads to delayed
action of the corresponding authorities and thus, the
burning of infested areas as the only viable solution
in accordance with the Mexican Official Standard
NOM-019-SEMARNAT-2017. This work proposes an
automated early detection and classification system for
bark beetles to streamline the monitoring process. The
system leverages Connected-Component Labeling for
the precise detection and counting of insects. For
classification, a modified ResNet50 residual neural
network is utilized, but with a modification to the custom
layer. Our approach achieves a high performance, with
an accuracy exceeding 90% on a dataset augmented
with over 3, 000 images. The system successfully
classifies two key species, Dendroctonus mexicanus
and Dendroctonus frontalis, demonstrating its potential
to significantly improve pest management efficiency
and reduce the need for drastic measures like
prescribed burns. This automated solution offers a
timely and effective alternative to traditional methods,
enabling a more proactive and targeted response to
forest infestations.