无人机遥感技术可以通过高分辨率的影像数据捕捉农田信息，及时、高效地识别并获取作物类型以用于农业统计调查。无人机影像蕴含遥感信息丰富，传统人工数字化或人机交互式农作物类型识别工作量大、效率低，影响无人机遥感即时性。深度学习作物分类模型往往需要海量、准确的训练数据。如何建立高精度、稳定、自动化的无人机农作物分类模型一直是遥感领域难点。

本文构建了完整的无人机水稻自动分类模型，首先研究了无人机影像耕地地块提取方法，基于SAM(Segment Anything Model, SAM)零样本分割模型设计完整后处理流程，在不同农业场景验证无人机影像耕地地块提取精度和效果。随后基于耕地地块提取结果进行面向地块的水稻作物分类，从光谱、纹理、空间角度尝试挖掘可量化的水稻地块特征，并组成特征规则集用于机器学习分类算法，同时还设计了自主学习地块特征的深度学习算法，分析比较这些模型在不同农业场景的分类精度和效果。得出如下结论：

（1）在耕地地块提取方面，SAM易受农作物种植结构影响。本文通过对比实验发现：多尺度分割在无人机影像中难以控制耕地地块的分割尺度，容易出现混合图斑问题；SAM零样本分割在各种农业场景下都能够完整分割耕地地块，模型受地块细碎化程度影响小，受农作物种植结构影响大，当影像中农作物种植结构复杂、区域异质性强时，SAM会对非耕地部分过分割，同时“双边界”、“孔洞”效应增强，整体耕地地块提取精度稳定在83.56%；经过后处理的SAM可以很好地保留SAM均质地块提取完整性，筛选了形状规则、有利于后续分类的耕地地块，在各种农业场景下均保证分割精度高于SAM，总体分割精度达86.17%。

（2）在耕地地块分类方面，基于特征的机器学习模型具有可解释、可视化的优点，SAM-UNet则具有更好的稳定性和适应性，但分类精度受农作物种植结构影响较大。本研究挖掘的基于光谱、纹理、空间形状的统计特征量具有很好的特征稳定性，以此作为输入特征训练机器学习分类器，结果证明HSL(Hue Saturation Lightness, HSL )特征和GLCM(Grey Level Co-occurrence Matrix, GLCM)特征对水稻的整体可分性优秀，独立分类时分别达到68.15%和74.14%的总体精度，LBP(Local Binary Pattern, LBP)特征对双季晚稻可分性较好，本文提出的垄状指数特征更适用于中稻的分类。本研究提出了基于SAM的水稻分类模型SAM-UNet，通过结合像素级语义信息和地块级宏观尺度信息实现水稻自动化分类。同时开展了对比实验分析各种农业场景下基于特征的机器学习方法和SAM-UNet的水稻分类精度和效果，结果表明，基于特征的机器学习模型易受图斑尺度和农作物种植结构的影响，分割结果不稳定，整体分类精度最高可达74.34%;SAM-UNet在各种农业场景下均能将晚稻和中稻地块提取准确提取，受图斑尺度影响小，受农作物种植结构影响大，复杂种植结构可能会使其总体精度下降约5.78%，最终分类总体分类精度稳定在81.23%。

Unmanned aerial vehicle (UAV) remote sensing technology can capture agricultural field information through high-resolution image data, enabling timely and efficient identification and acquisition of crop types for agricultural statistical surveys. UAV imagery contains rich remote sensing information. Traditional manual digitization or human-machine interactive crop type identification involves large workloads and low efficiency, affecting the timeliness of UAV remote sensing. Deep learning crop classification models often require massive and accurate training data, resulting in high costs for UAV image dataset creation. Establishing a high-precision, stable, and automated UAV crop classification model has long been a challenge in the remote sensing field.

In response to the above issues, this study conducted comparative experiments on UAV image field extraction methods, designed a complete post-processing workflow based on Segment Anything Model (SAM) zero-shot segmentation, and utilized SAM segmentation results to guide UNet for fine-grained semantic segmentation to extract target agricultural field patches. The extraction effectiveness of agricultural field patches from different agricultural landscapes using different models was verified. Subsequently, rice crop classification based on field patch extraction results was carried out. Quantifiable rice field features were explored from spectral, textural, and spatial perspectives, and feature separability analysis was performed to compose a feature rule set for machine learning classification algorithms. Additionally, deep learning algorithms for autonomously learning field features were designed and compared, analyzing the classification accuracy and effectiveness of these features in different agricultural landscapes. The following conclusions were drawn:

(1) Regarding the extraction of cultivated land parcels, SAM is easily influenced by crop planting structures. Comparative experiments revealed the following: Multi-scale segmentation in UAV imagery struggles to control the segmentation scale of cropland parcels, leading to the issue of mixed patches. In contrast, SAM zero-shot segmentation can completely segment cropland parcels across various agricultural landscapes, with the model being less affected by the fragmentation of parcels but significantly influenced by crop planting structures. When crop planting structures are complex and regional heterogeneity is high, SAM tends to over-segment non-cropland areas, while the "double-boundary" and "hole" effects become more pronounced. The overall cropland parcel extraction accuracy remains stable at 83.56%. Post-processed SAM effectively retains the integrity of SAM’s homogenous parcel extraction, filtering out regularly shaped cropland parcels that are beneficial for subsequent classification. This approach ensures segmentation accuracy higher than that of SAM across various agricultural landscapes, achieving an overall segmentation accuracy of 86.17%.

(2) In terms of cultivated land parcel classification, Feature-based machine learning models have the advantages of interpretability and visualization, while SAM-UNet demonstrates better stability and adaptability. However, its classification accuracy is significantly affected by crop planting structures. This study explores the statistical features based on spectral, textural, and spatial shape characteristics, which exhibit good feature stability. These features were used as input to train machine learning classifiers. The results show that HSL (Hue Saturation Lightness) and GLCM (Grey Level Co-occurrence Matrix) features provide excellent separability for rice, achieving overall accuracies of 68.15% and 74.14% respectively when classified independently. LBP (Local Binary Pattern) features offer better separability for late-season double-cropping rice, while the proposed ridge index feature is more suitable for the classification of medium-season rice. This study proposes a SAM-based rice classification model, SAM-UNet, which achieves automated rice classification by combining pixel-level semantic information with parcel-level macro-scale information. Comparative experiments were conducted to analyze the classification accuracy and performance of feature-based machine learning methods and SAM-UNet under various agricultural landscapes. The results indicate that feature-based machine learning models are susceptible to the influences of patch scale and crop planting structures, leading to unstable segmentation results with a maximum overall classification accuracy of 74.34%. On the other hand, SAM-UNet accurately extracts late-season and medium-season rice parcels across different agricultural landscapes, being less affected by patch scale but significantly influenced by crop planting structures. Complex planting structures can reduce its overall accuracy by approximately 5.78%, with the final overall classification accuracy stabilizing at 81.23%.