As the unmanned aerial vehicles (UAVs) continue to be deployed for various mission-critical data acquisition, localized computing on the drone-acquired data for efficient analysis, without significantly impacting the limited resources on board the drone, has emerged as a formidable research challenge. In this article, we address this issue with a natural resource management use-case whereby early forest-fire detection using the popular convolutional neural network (CNN)-based inference models are considered in the drone, which leads to resource exhaustion. To alleviate this, we propose a lightweight hierarchical artificial intelligence (AI) framework, which adaptively switches between a simple machine learning-based model and an advanced deep learning-based CNN model. Then, we formulate a multi-objective optimization problem to model the tradeoff between forestfire detection accuracy and computational performance. We obtain the Pareto-optimal solution of the formulated problem by optimizing a new hyperparameter (i.e., the confidence score threshold) by employing the technique for order of preference by similarity to ideal solution (TOPSIS) for the whole model. Thus, we alleviate the computational burden while retaining a high level of detection accuracy. Finally, based on a real dataset, empirical results are reported to evaluate the performance of our proposal in terms of its lightweight features. IEEE