Scholarlink

<p>Human–elephant conflict (HEC) remains a serious ecological and societal encounter, resulting in loss of human lives, crop destruction, and elephant mortality. Early detection of elephants and an accurate understanding of their behavioral state are essential for effective conflict mitigation. This work proposes a hybrid deep learning scheme for elephant detection (ED) and State-of-Mind (SoM) recognition, integrating handcrafted acoustic features and deep visual–spectral representations. Two complementary approaches are employed: (i) conventional audio feature extraction using Mel-frequency cepstral coefficients (MFCC), spectrogram, spectral centroid, and roll-off features followed by preprocessing and deep neural network (DNN) classification, and (ii) mel-spectrogram image representation processed using a pre-trained Visual Geometry Group-16 (VGG-16) network via transfer learning. To further enhance prevention, a Q-Learning–based decision-making module is introduced to dynamically trigger deterrent actions based on detected elephant presence and inferred SoM. The proposed system enables proactive, intelligent prevention of elephant–human encounters by combining perception, behavioral analysis, and reinforcement learning–based control. Experimental results demonstrate improved detection accuracy (99.3%), robust SoM classification (94.1%), and effective adaptive response strategies for controlled deployment in forest boundary regions.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Elephant detection and behavioral state analysis using hybrid acoustic deep learning and reinforcement learning

  • 期刊论文

  • 发表日期：

    2026年4月29日

查看全文
  • V. Karthikeyan,
  • R. Varun Prakash

摘要

Human–elephant conflict (HEC) remains a serious ecological and societal encounter, resulting in loss of human lives, crop destruction, and elephant mortality. Early detection of elephants and an accurate understanding of their behavioral state are essential for effective conflict mitigation. This work proposes a hybrid deep learning scheme for elephant detection (ED) and State-of-Mind (SoM) recognition, integrating handcrafted acoustic features and deep visual–spectral representations. Two complementary approaches are employed: (i) conventional audio feature extraction using Mel-frequency cepstral coefficients (MFCC), spectrogram, spectral centroid, and roll-off features followed by preprocessing and deep neural network (DNN) classification, and (ii) mel-spectrogram image representation processed using a pre-trained Visual Geometry Group-16 (VGG-16) network via transfer learning. To further enhance prevention, a Q-Learning–based decision-making module is introduced to dynamically trigger deterrent actions based on detected elephant presence and inferred SoM. The proposed system enables proactive, intelligent prevention of elephant–human encounters by combining perception, behavioral analysis, and reinforcement learning–based control. Experimental results demonstrate improved detection accuracy (99.3%), robust SoM classification (94.1%), and effective adaptive response strategies for controlled deployment in forest boundary regions.