Smoke Recognition in Satellite Imagery via an Attention Pyramid Network With Bidirectional Multilevel Multigranularity Feature Aggregation and Gated Fusion

Recognizing smoke in satellite imagery is a critical approach in an Internet of Things (IoT) system for monitoring forest fires. However, the task remains challenging due to false alarms of smoke-like occurrences caused by complex land cover types, and missing detections caused by the diversity of fire smoke. Some reasons are that existing methods overlook attention granularity, neglect all-layer-based fusion of low-level features with high-level semantic information, and fail to address interferences arising from fusing different kinds of features. To solve these issues, this article presents an attention pyramid network with bidirectional multilevel multigranularity feature aggregation and gated fusion for smoke recognition. First, to guide the model sequentially extract multigranularity smoke attention clues for complementary smoke perception, we design an attention-guided feature pyramid module by concatenating residual blocks and attention pyramid blocks. Second, to leverage both low-level fine-grained and high-level semantic features in all network layers, we design a bidirectional feature aggregation module using multilevel multigranularity feature blocks. Finally, to selectively integrate the features with different resolutions and semantic levels to effectively achieve feature complementarity and avoid feature mutual interference, we design a gated feature fusion module using gated feature fusion blocks. The experimental results demonstrate that our model achieves an accuracy of 98.33% on the USTC-SmokeRS data set. Additionally, on the E-USTC-SmokeRS data set, our model achieves a detection rate of 94.92%, a false alarm rate of 3.00%, and an F1-score of 0.9553. These results surpass the performance of existing satellite-imagery-based smoke recognition methods.

I. Introduction

Wildfire is a kind of natural disaster that is sudden-happening, destructive, and harmful to the ecosystems, human lives, and property, leading to local atmospheric pollution and the loss of plant resources [1]. Smoke always occurs before the wildfire and it is the first reliable sign in case of a wildfire. Thus, recognizing smoke is effective to find wildfires as quickly as possible to avoid the expansion of fire and minimize the damage to the ecological systems, infrastructure, and human beings. In current smoke recognition technology [2], [3], forest surveillance videos are initially captured via high-resolution cameras installed atop hills or mobile towers. Smoke recognition methods are then applied to analyze these videos and identify smoke automatically. As each camera has a limited monitoring view of the forest, it is usually necessary to deploy multiple cameras to cover a vast area.