Lepidopterans, which include butterflies and moths, have been studied extensively due to their ecological significance being sensitive indicators of environmental change. Monitoring these species helps scientists assess the health of ecosystems, track changes in biodiversity, and understand the impacts of climate change, pollution, and habitat loss. Additionally, studying Lepidopterans can contribute to conservation efforts and agricultural management, as some species are pollinators while others are pests. In this study, we examined two different methods for lepidopterans detection and classification using data from iNaturalist. Object detection using neural networks relies on high-quality image labels or box bounds; thus, the task of using large, unlabeled images from the online image repository iNaturalist with a traditional neural network proves to be tedious. Therefore, we propose an approach that uses both zero-shot and few-shot learning for Lepidopterans detection. For the zero-shot model, Detic model was adapted and tested. The primary advantage of this method is that traditional neural-network detections require tediously achieved, well-labeled datasets, which can be traded for an algorithm that couples a natural language model (to classify boxes) with Detic (to identify boxes). For our few-shot model, we custom-labeled our images using MakeSense and built our model using Detectron2’s segmentation model and a Mask R-CNN. Our data has shown that the zero-shot detection of the Detic model performs with a higher accuracy than our few-shot Detectron2 model. Additionally, we have hand-labeled a dataset of 4000 images of moths, butterflies, caterpillars, and cocoons that can be used to test other supervised learning models.