16:15 〜 16:30
[HCG26-09] Cross-domain Transfer application with Deep Learning-based segmentation to detect Agricultural Burned Areas: A Case Study in Punjab, India
キーワード：Transfer Learning, Emission Inventory, Agricultural Burn Area, Sentinel-2
Since the advent of the Green Revolution in India and availability of electricity at the agricultural fields in the mid 1980s, many Farmers in North India use mechanized harvesters, leaving the crop residue scattered on the ground. This residue is often disposed of through open burning as a quick method to prepare the land for the next seasonal crop. This raises the question of, to what extent this emission particularly PM2.5, from burning contributes to the already hazardous air quality in this region, especially the capital New Delhi. Global fire emission inventory extensively utilizes the Moderate Resolution Imaging Spectroradiometer (MODIS) on NASA's Terra and Aqua satellites for burn area data, the relatively coarse spatial resolution of MODIS (ranging from 500 meters to 1 kilometer) often results in the underestimation of fire events.
In our study, we have implemented a deep learning-based segmentation model combined with Hi-Resolution Sentinel-2 imagery and ground-level field observation as inputs to improve the fire estimates over the state of Punjab. The model was first trained over Portugal using Sentinel-2 data and reference data for forest fire, particularly focussing on months with the highest number of fire events, from the ICNF (Portuguese Institute for Nature Conservation and Forests). While the model demonstrated promising results in Portugal, with few inaccuracies, the initial transfer application of the model in Punjab encountered challenges due to significant differences in forest fires and crop residue burning. This resulted in overestimations of burn areas and misidentifications in urban and aquatic regions. Addressing the need for improved model adaptation to the Punjab environment, we conducted manual annotation of over 100 square kilometers of land in Punjab, ensuring the exclusion of water bodies and the verification of marked burned areas against Google Earth imagery, as well as MODIS and VIIRS fire hotspot data. We also utilized geolocated burned area images captured during a field observation campaign in Punjab, as a reference for model evaluation.
The pre-trained model utilizes annotated data and corresponding Sentinel-2 spectral bands B03 (green), B8A (near-infrared), and B11 (short-wave infrared) to identify burn areas with greater precision. We assessed the model's accuracy through the Dice coefficient and Intersection over Union (IOU) metrics. To benchmark our model's performance, we compared it against a baseline model that calculates the Normalized Burn Ratio (NBR) using near-infrared and short-wave infrared bands. Initially, the pre-trained model showed modest effectiveness in Punjab, with an average IOU of 0.033 and a Dice score of 0.061. However, after fine-tuning, the model demonstrated a significant improvement, achieving an average IOU of 0.44 and a Dice score of 0.60, thereby markedly surpassing both the pre-trained and baseline models in accuracy.
The primary goal of this work is to assess the potential of a deep learning (DL) model, trained on regions with abundant ground truth data, to analyze areas with scarce or absent ground truth data. This substantial enhancement in model performance highlights the critical role of tailored approaches in the precise detection of burn areas.
In our study, we have implemented a deep learning-based segmentation model combined with Hi-Resolution Sentinel-2 imagery and ground-level field observation as inputs to improve the fire estimates over the state of Punjab. The model was first trained over Portugal using Sentinel-2 data and reference data for forest fire, particularly focussing on months with the highest number of fire events, from the ICNF (Portuguese Institute for Nature Conservation and Forests). While the model demonstrated promising results in Portugal, with few inaccuracies, the initial transfer application of the model in Punjab encountered challenges due to significant differences in forest fires and crop residue burning. This resulted in overestimations of burn areas and misidentifications in urban and aquatic regions. Addressing the need for improved model adaptation to the Punjab environment, we conducted manual annotation of over 100 square kilometers of land in Punjab, ensuring the exclusion of water bodies and the verification of marked burned areas against Google Earth imagery, as well as MODIS and VIIRS fire hotspot data. We also utilized geolocated burned area images captured during a field observation campaign in Punjab, as a reference for model evaluation.
The pre-trained model utilizes annotated data and corresponding Sentinel-2 spectral bands B03 (green), B8A (near-infrared), and B11 (short-wave infrared) to identify burn areas with greater precision. We assessed the model's accuracy through the Dice coefficient and Intersection over Union (IOU) metrics. To benchmark our model's performance, we compared it against a baseline model that calculates the Normalized Burn Ratio (NBR) using near-infrared and short-wave infrared bands. Initially, the pre-trained model showed modest effectiveness in Punjab, with an average IOU of 0.033 and a Dice score of 0.061. However, after fine-tuning, the model demonstrated a significant improvement, achieving an average IOU of 0.44 and a Dice score of 0.60, thereby markedly surpassing both the pre-trained and baseline models in accuracy.
The primary goal of this work is to assess the potential of a deep learning (DL) model, trained on regions with abundant ground truth data, to analyze areas with scarce or absent ground truth data. This substantial enhancement in model performance highlights the critical role of tailored approaches in the precise detection of burn areas.