基于FIDT图的稠密小目标树木检测架构

Dense and Small Tree Detection Framework Based on FIDT Map

摘要:

在高密度区域，树木冠层重叠和遮挡严重，传统方法难以准确识别单棵树木，影响成活率和绿化覆盖率的评估精度，因此，亟需一种高精度和强鲁棒性的检测方法，以提升城市树木成活监测的准确性和效率。提出基于焦距反比变换图通过计算目标点到最近边界的距离，对稠密树木小目标检测方法，采用一种多层次的局部极大值检测策略，能够有效提取树木目标中心点信息，并区分相邻目标，从而降低重叠区域目标检测误差。此外，为了增强模型对局部结构信息的学习能力，引入独立结构性损失。选用一个高分辨率林地遥感数据集，并对稠密小目标树木进行检测与定位分析。为了验证本文方法的有效性，与多种现有深度学习方法进行对比分析，评估不同方法在精确率、召回率、定位误差等方面表现。试验结果表明，本文方法在城市林地遥感数据集上展现了较高的定位和计数性能。在计数任务中，本文方法平均绝对误差（MAE）和均方误差（MSE）分别为7.87和10.23，相较于现有方法CSRNET分别降低45.1%和48.3%；在定位任务中，本文方法在Claremont、Long_beach、Palm_springs和Riverside 4个区域的F1值分别达到75.2%、72.7%、74.8%和71.5%，较现有基于密度图法至少提升19个百分点，同时在精确率和召回率等指标上也实现了全面超越。研究结果为林业资源管理、生态监测和环境保护提供了一种高效、可靠的技术手段。

Abstract:

In high-density areas, the severe overlap and occlusion of tree canopies pose significant challenges for traditional methods in accurately identifying individual trees, thereby compromising the precision of evaluations related to survival rates and greening coverage. Consequently, there is an urgent need for a detection approach with high accuracy and strong robustness to enhance the effectiveness and reliability of urban tree survival monitoring. The detection method proposed was based on the focal inverse distance transform (FIDT), which computed the distance from each target point to its nearest boundary and performed an inverse transformation. On this basis, a multi-level local maxima detection strategy was introduced to effectively extract the center point information of tree targets and distinguish adjacent targets, thereby reducing detection errors in overlapping regions. Additionally, an independent structural loss was incorporated to enhance the model’s ability to learn local structural information. Experiments were conducted on a high-resolution forest remote sensing dataset to detect and localize dense small tree targets. To validate the effectiveness of the proposed method, comparative analyses were performed against various existing deep learning approaches, evaluating their performance in terms of detection accuracy, recall, and localization error. Experimental results showed that the proposed method achieved strong localization and counting performance on the urban forest dataset. For counting, it attained MAE and MSE of 7.87 and 10.23, reducing errors by 45.1% and 48.3% compared with that of CSRNET. For localization, F1-scores in Claremont, Long_beach, Palm_springs, and Riverside were 75.2%, 72.7%, 74.8%, and 71.5%, averaging over 19 percentage points higher than that of existing density map-based methods, with overall improvements also observed in precision and recall. The research result can provide an efficient and reliable technical approach for forestry resource management, ecological monitoring, and environmental protection. In the future, integrating multi-source remote sensing data and time series analysis could further improve the accuracy and efficiency of dynamic forest resource monitoring.