松类造型苗模型训练样本数据

1. 精准农业与管理：通过对作物冠层形态、颜色的精细标注，训练的模型可用于监测作物长势、精准估算产量、早期发现病虫害及杂草分布，从而指导变量施肥、施药，大幅提升农业生产效率和资源利用率。
2. 生态环境监测与保护：利用标注好的样本数据，模型可学会识别不同树种、草地乃至入侵物种。其应用包括森林资源调查、生物多样性评估、生态环境破坏（如非法砍伐）监测以及湿地、自然保护区生态变化的动态跟踪与保护。
3. 林业调查与碳汇核算：模型能够对林木种类、健康状态、郁闭度甚至单木进行识别和统计，应用于森林蓄积量估算、林业病虫害预警、火灾后受损评估等，为林业碳汇的精准计量提供重要数据支撑。
4. 灾害应急与评估：在洪涝、山体滑坡、森林火灾等自然灾害发生后，通过训练模型快速识别受灾区域的植被损毁情况，为灾情评估、救援部署及灾后生态恢复规划提供及时、准确的决策依据。
高质量的训练样本数据是植被识别模型在实际场景中发挥效用的基石，使得低空遥感技术能够从“看得见”升级为“看得懂”，赋能多个行业的数字化、智能化管理。1、数据采集：利用飞马v500无人机，利用自动拼图得到1cm-3cm分辨率的无人机正射影像数据，并设置CGCS2000 / 3°投影坐标系，影像分辨率以及影像坐标系等参数同步加入至影像数据中。
2、数据预处理以及数据标注：首先，选择适用于样本的影像，明确裁切区域并绘制范围矢量；利用矢量数据裁切影像，采用在线标记点位并添加作物属性的方式，以点位为中心点进行裁剪，生成指定尺寸（如256*256）的影像切片。
3、数据集设置以及模型选择：按照7:2:1的比例设置训练集、验证集和测试集。使用自行搭建的TransCNN-Vision模型进行训练。
5、训练设置：模型选择与初始化以vision_transformer的large模型为权重文件，初始化常规模型参数后，读取数据集文件夹个数确定模型分类数，最优学习率为0.001，同时冻结出head和pre_logits外的所有权重参数，batchsize，根据当前设备现存余量自动调整，默认值为16，根据样本分辨率动态调整patch_size用于提高不同分辨率下的特征捕获能力。最后利用自适应高精度模型保存策略，自动保存训练精度mDice（Mean Dice Coefficient）指标最高模型的模型作为最佳模型。mDice计算公式如下： mDice=2*|X∩Y|/(|X|+|Y|)。
训练mDice指数基于对测试集样本数据预测并计算获得，其中|X∩Y|为预测结果与真实标注的交集，| X |和| Y |分别为预测结果与真实标签的各自的数量之和。
5、模型精度评估：通过在真实影像中进行模型提取并人工校正，实现对模型在真实场景中的提取效果。利用提取错误率和提取遗漏率指标来评估被识别物模型的提取能力，提取错误率用于评估模型提取结果中不是合理的比例，提取遗漏率用于评估模型提取结果依然没有提取出被识别物的比例，提取遗漏率越接近4.2%，提取错误率越接近6.2%，表明当前提取结果准确率越高，能够降低的成本越高。提取错误率计算公式如下：（|X|-|X∩Y|）/|X|。
提取遗漏率计算公式如下：（|Y|-|X∩Y|）/|Y|。其中，|X∩Y|为正确识别为被识别物的数量，| X |和| Y |分别为预测和真实的被识别数量。考虑到真实场景的复杂性，提取错误率和提取遗漏率保持在10%以内即可视为结果具有较高的准确性。

1. Precision Agriculture and Management: With fine annotations of crop canopy morphology and color, the trained model can be used to monitor crop growth, accurately estimate yield, detect pests, diseases and weed distribution at an early stage, so as to guide variable-rate fertilization and pesticide application, and greatly improve agricultural production efficiency and resource utilization rate.
2. Ecological Environment Monitoring and Protection: Using annotated sample data, the model can learn to identify different tree species, grasslands and even invasive species. Its applications include forest resource surveys, biodiversity assessments, monitoring of ecological environment damage (such as illegal deforestation), and dynamic tracking and protection of ecological changes in wetlands and nature reserves.
3. Forestry Survey and Carbon Sink Accounting: The model can identify and count tree species, health status, canopy density and even individual trees. Its applications include forest stock volume estimation, forestry pest and disease early warning, post-fire damage assessment, etc., providing important data support for accurate measurement of forestry carbon sinks.
4. Disaster Emergency Response and Assessment: After natural disasters such as floods, landslides and forest fires, the trained model can quickly identify the vegetation damage in the affected areas, providing timely and accurate decision-making basis for disaster assessment, rescue deployment and post-disaster ecological restoration planning.

High-quality training sample data is the cornerstone for vegetation recognition models to exert their effects in real-world scenarios, enabling low-altitude remote sensing technology to upgrade from "being able to see" to "being able to understand", and empowering digital and intelligent management of multiple industries.

1. Data Collection: Using the FEIMA V500 unmanned aerial vehicle (UAV), UAV orthophotograph data with a resolution of 1 cm to 3 cm is obtained via automatic stitching. The China Geodetic Coordinate System 2000 / 3° projection coordinate system is set, and parameters such as image resolution and image coordinate system are synchronously added to the image data.
2. Data Preprocessing and Annotation: First, select images suitable for samples, define the cropping area and draw the range vector; crop the image using the vector data, mark points online and add crop attributes, then crop with the point as the center to generate image slices of specified dimensions (such as 256*256).
3. Dataset Setup and Model Selection: The training set, validation set and test set are set at a ratio of 7:2:1. Training is carried out using the self-built TransCNN-Vision model.
5. Training Setup: The model selection and initialization use the vision_transformer large model as the weight file. After initializing the conventional model parameters, the number of model classification categories is determined by reading the number of dataset folders. The optimal learning rate is 0.001, and all weight parameters except head and pre_logits are frozen. The batch size is automatically adjusted according to the remaining memory of the current device, with a default value of 16. The patch size is dynamically adjusted according to the sample resolution to improve feature capture capability under different resolutions. Finally, the adaptive high-precision model saving strategy is used to automatically save the model with the highest training accuracy metric mDice (Mean Dice Coefficient) as the best model.

The formula for mDice is as follows: $mDice=2 imes|Xcap Y|/(|X|+|Y|)$

The training mDice metric is obtained by predicting and calculating the test set sample data, where $|Xcap Y|$ is the intersection of the prediction results and the ground-truth annotations, and $|X|$ and $|Y|$ are the total number of predicted results and ground-truth labels respectively.

5. Model Accuracy Evaluation: The extraction effect of the model in real-world scenarios is evaluated by extracting from real images and performing manual correction. The extraction error rate and extraction omission rate are used to evaluate the extraction capability of the recognition model. The extraction error rate is used to evaluate the proportion of unreasonable results in the model's extraction outputs, while the extraction omission rate is used to evaluate the proportion of targets that are not extracted by the model. The closer the extraction omission rate is to 4.2% and the extraction error rate is to 6.2%, the higher the accuracy of the current extraction results and the more cost can be reduced.

The formula for extraction error rate is: $ ext{Extraction Error Rate}=(|X|-|Xcap Y|)/|X|$

The formula for extraction omission rate is: $ ext{Extraction Omission Rate}=(|Y|-|Xcap Y|)/|Y|$. Here, $|Xcap Y|$ is the number of targets correctly recognized, and $|X|$ and $|Y|$ are the number of predicted and ground-truth targets respectively. Considering the complexity of real-world scenarios, if the extraction error rate and extraction omission rate are both within 10%, the results can be considered to have high accuracy.

搜集汇总

背景与挑战

背景概述

该数据集是松类造型苗模型训练样本数据，包含2459条xlsx格式记录，涵盖图像ID、分辨率、采集时间等结构化字段，专用于训练植被识别模型。数据通过无人机采集，应用场景包括精准农业和林业调查，模型提取错误率和遗漏率分别为4.2%和6.7%，具有高准确性和多行业实用性。