In Kenya, conflicts between humans and elephants often lead to destruction of crops, lower family income, and put people and elephants at risk at the border of developing farms. Fences, patrols, and manual camera inspection are all expensive and take too long to be useful. This research creates an affordable early warning system that uses transfer learning with convolutional neural networks to find elephants in images. There are two label of elephant and non-elephant wildlife which are a public wildlife corpus of approximately 40,000 photographs of binary task. Cleaning of the dataset was done to ensure high quality and consistency. To deal with class skew, we oversampled elephants in the training split while keeping the natural imbalance in the validation and test sets to reflect real-world conditions. Five ImageNet-pretrained backbones (VGG16, ResNet50, InceptionV3, Xception, MobileNetV2) are tested using a single two-phase regimen: first, a new binary head is trained with the base frozen, and then the upper layers are fine-tuned using early stopping and learning-rate scheduling. Performance is provided using accuracy, precision, recall, F1, ROC-AUC, PR-AUC, and confusion matrices. The recall-first threshold is set to minimize missing elephants. All transfer models have almost ideal ranking quality (ROC-AUC=1.00, PRAUC > 0.99). When the recall-first operating point is reached, recall becomes close to 1.0 and accuracy stays around 0.56–0.59, which is a good balance for safety-critical warnings. Xception has the best overall separation and is exported as the main service model. MobileNetV2 has the same recall but reduced latency for edge cases. A FastAPI service standardizes inputs, sets calibrated thresholds, makes warnings that can be audited, and adds Grad-CAM overlays for transparency. It gives early-warning information that can be used to help coexist and preserve crops.
Pascaline Jerotich (Wed,) studied this question.