Monitoring the Mangrove Species in Hong Kong with High Resolution Images Using Deep Learning Networks

Mangrove forest is a highly productive tropical and subtropical ecosystem, providing significant ecological functions and valuable social-economic services, including coastal protection, aquaculture, habitat provision, biodiversity protection, and carbon sequestration. Due to coastal development, mangroves have been suffering a significant loss in extent as well as species. Mangrove species richness has profound influences on the mangrove system in terms of robustness and health, but it cannot be restored by simply introducing exotic species, which may be invasive and threatening to local species and the ecosystem. Therefore, monitoring mangroves at the species level is necessary and useful for mangrove conservation. Remote sensing provides a cost-effective and feasible way to observe mangrove forests. The availability of high-resolution images makes monitoring fragmentized mangrove patches at the species level from regionality to globality possible. However, accurate mapping of mangrove species remains a challenge due to morphological and spectral similarity between different species and large inter-class variability within each species. This research aims to systematically explore the spatial and spectral information provided by high spatial resolution multispectral and hyperspectral satellite images for mangrove species mapping to facilitate mangrove management.First, the conventional methods of extracting spatial features, grey level co-occurrence matrix (GLCM) and Gaussian Markov Random Field (GMRF), are analyzed and compared, demonstrating the difficulties in feature design for mangroves without distinguished shapes, edges and texture. Instead of feature engineering, deep learning driven by a large amount of data is introduced to extract mangroves' abstract features (high-level features). For patched mangroves, a small-patched convolutional neural network (CNN) is proposed to overcome the limitation of fixed large inputs for conventional CNNs. The proposed method obtains higher accuracy with modification in pooling layers, small kernels, and lightweight architecture. The results show that the abstract features from deep learning can improve the mapping of all mangrove species in the study sites, and the contextual information rather than mangroves per se making more contribution.Second, spectral features from hyperspectral images are also explored for mangrove species mapping, aiming to answer whether spectral increment can improve the performance of mangrove species mapping. Based on true Hyperion, the Hyperion with the same period as Gaofen 5 hyperspectral images and Landsat 8 was simulated to form a series of spectral features with different resolutions. The comparison using support vector machine and Random Forests demonstrates that significant improvement can be obtained from hyperspectral features over multispectral features (73.89% vs. 86.82%). A further improvement in spectral resolution from 10nm to 5nm brings limited effects (87.12% vs. 86.82%), in which the visible/near-infrared spectrum makes a significant contribution to better mapping, particularly mapping AC species.Third, to fully leverage the high spatial and high spectral information from multispectral and hyperspectral images for mangrove species mapping, a novel scheme is proposed to account for spectral variability between two images, and a lightweight 3D CNN was proposed to explore the fusion of spatial and spectral information. The synthetic high-resolution hyperspectral image can improve the problem of limited samples for deep learning and concurrence of high spatial and high spectral information, especially for species mapping of patched mangroves at the region level. Compared to hyperspectral and multispectral images alone, the results show that image fusion can greatly improve mangrove species mapping (91.25% vs. 73.05% and 85.52%), and 3D CNN excels SVM in processing high dimension data.Finally, a case study of individual invasive mangrove species of Sonneratia detection using RetinaNet based on CNN shows how to support mangrove management with high-resolution images using deep learning. Three thousand six hundred seventy-eight individuals were detected in Mai Po, and most of them are with the crown size of 5 by 5 meters. It can provide early warning for invasive species and support invader control.The primary outcome of this research provides a feasible way for accurate mapping of mangrove species in Hong Kong with up-to-date methods of deep learning. The method provides the baseline for extracting spatial and spectral information and fusing them for mangrove species mapping. The local government can get subtle information on mangroves timely and make quick responses to support effective and efficient conservation.紅樹林是生長在熱帶亞熱帶的一種高生產能力的植被，提供了重要的生態功能和社會經濟服務，包括海岸帶保護、水產業、為各種生物提供栖息地、生物多樣性保護、固碳儲碳等。由於地處人類活動密集的海岸帶，紅樹林面積正遭受著流失，其物種類別也在下降。紅樹林的品種多樣性會嚴重影響整個紅樹林生態的健康、健壯性及可持續性。然而，品種的下降是無法通過引進外來品種實施造林得以恢復。相反，外來品種還可能會帶來物種入侵風險，對本地物種構成新的威脅。因此，我們急需對紅樹林進行物種級別的監測以支持對紅樹林的高效管理。遙感提供了一種經濟、可行的觀測紅樹林的方法，尤其是高分辨率影像的日漸普及和易獲取性，使得對區域乃至全球範圍内的紅樹林斑塊進行物種分類成爲可能。然而，由於不同紅樹林品種在光譜和外在形態上存在相似性且同一品種存在較大的類内差異，精確的紅樹林種間分類是一項挑戰的任務。鑒於此，本論文通過系統地探索不同紅樹林品種在高分辨率影像（高空間分辨率多光譜和高光譜分辨率高光譜影像）上呈現的空間和光譜特徵，實現紅樹林種間分類，進而支持紅樹林管理。首先，對傳統空間特徵提取方法，灰度共生矩陣和馬爾可夫隨機模型，進行了分析和對比，並對紅樹林進行了種間分類，發現傳統方法很難描述沒有規則形狀、明顯邊界及清晰紋理紅樹林的空間特徵。深度學習，不同與特徵工程人爲構建特徵描述，它通過數據驅動，可以從大量數據從學習紅樹林的高級抽象特徵。爲了適用於紅樹林，提出了一種小斑塊深度捲積網絡，剋服了傳統捲積網絡對固定大尺寸圖像輸入的需求。通過修改深度網絡池化層、應用小核和輕網絡架構，提出的方法可以獲得更高的精度。結果表明，深度網絡提取的抽象特徵可以有效地提高實驗區内每個物種的識別精度；其中，上下文信息提供了比物種本身更多的信息。其次，對高光譜數據的光譜特徵影響紅樹林種間分類進行了探索，旨在回答光譜分辨率的提高是否有助於紅樹林種間分類的問題。基於真實Hyperion，通過模擬與高分5號高光譜和Landsat 8多光譜同時期的Hyperion數據，形成不同光譜分辨率的數據，結合傳統的支持向量機和隨機森林兩種可處理高維特徵的機器學習方法，對比了不同分辨率光譜特徵對紅樹林種間的影響。結果顯示，相比多光譜特徵，高光譜特徵可以極大地提高紅樹林中間分類（73.89% vs 86.82%）,光譜分辨率從10納米進一步提高到5納米，對紅樹林種間分類提升有限（86.82% vs. 87.12%）,主要的提升在於對桐花樹物種的識別；主要的貢獻來自於可見光和近紅外光譜分辨率的提高。此外，爲了更好地利用多光譜影像的空間信息和高光譜影像的光譜信息對不同紅樹林品種進行分類，本文在考慮兩種影像光譜差異情況下提出一個集成數據融合和輕3D深度捲積網絡的新模式實現空譜特徵的融合。多光譜和高光譜數據的融合可以在保持空間和光譜信息的同時解決深度學習對多樣本的需求，特別適用於區域紅樹林斑塊的種間分類。對比單獨使用高光譜和多光譜數據（73.05%和85.52%），結果表明融合的數據可以提高紅樹林中間分類精度（91.25%），3D深度捲積網絡在處理高維數據時比支持向量機更有優勢。最後，本文結合香港米埔保護區外來物種海桑管理的需求，展示了使用高分辨率影像結合深度學習（RetinaNet）實現單株海桑的檢測，進而支持紅樹林的管理。在實驗區共檢測出3687株海桑，且估算出大部分海桑的冠幅大小為5米x5米。這些信息有助於對入侵物種的早期預警及移除工作。本研究提供了一種可行的基於深度學習的高精度紅樹林種間分類的方法，包括如何提取紅樹林的空間和光譜特徵，以及如何融合空間和光譜特徵服務于紅樹林種閒分類。紅樹林種間分類可以為政府提供更精確的信息，以便采取及時有效的措施，支持高效的保護和管理。"WAN, Luoma.Ph.D. Chinese University of Hong Kong 2021.Includes bibliographical references (leaves )Abstracts in English and ChineseTitle from PDF title page (viewed on …