Deep Learning-Driven Transformation of Remote Sensing Education for Ecological Civilization and Sustainable Development

Abstract

Against the background of China’s ecological civilization construction and sustainable development strategies, how remote sensing courses adapt to the demands of the artificial intelligence era has become an urgent issue for undergraduate education in relevant disciplines at universities. This study proposed a trinity teaching reform path of “deep learning and remote sensing, and ecological sustainability”, aiming to cultivate interdisciplinary talents with capabilities in intelligent interpretation and practical application. The study established a three-stage curriculum objective system, integrating knowledge, ability, and literacy, designed a five-dimensional linkage teaching method combining case-driven teaching, modular training, and blended learning, and conducted teaching practices using mainstream deep learning frameworks and cloud platforms. Through hierarchical teaching practice cases and multi-dimensional evaluation data, it was shown that the reform effectively enhanced the experiment group students’ abilities in deep learning applications, complex remote sensing data processing, and ecological problem-solving. The achievement values for all five evaluation indicators exceeded 80%, with the highest improvement reaching 28% compared to the control group. The results indicate that this teaching reform not only enhances learning outcomes but also provides a valuable framework and practical pathway for remote sensing education empowered by artificial intelligence and the cultivation of professional talent in future sustainable development fields.

1. Introduction

With the increasingly severe global environmental problems, ecological civilization construction has emerged as a central concern worldwide and a critical pathway toward achieving sustainable development. In China, strategic measures, such as the “dual carbon” goal, the red line of ecological conservation, and the biodiversity monitoring programs, need more advanced ecological environment monitoring technology. Remote sensing technology is an important means to obtain earth surface information. It plays an essential role in ecological parameter inversion, real-time environmental monitoring, and ecological risk assessment [1,2,3]. It has become a crucial tool to support sustainable development decision-making [4,5]. The cultivation of highly qualified remote sensing professionals for ecological environment applications is closely linked to the achievement of ecological civilization and Sustainable Development Goals. This, in turn, imposes new requirements on university remote sensing courses, demanding that their teaching contents and methods be intelligent, practical, and frontier [6].

However, the teaching of traditional remote sensing courses faces significant challenges [7,8,9]. The content of these courses mainly focuses on fundamental theories and conventional digital image processing methods. Cutting-edge capabilities are insufficiently covered. These include the intelligent interpretation of massive multi-source remote sensing data, in-depth recognition of complex feature patterns, and time-sequence modeling of dynamic ecological processes [10]. For instance, there is an excessive focus on spectral physical mechanisms and classical digital image processing (e.g., NDVI calculation, maximum likelihood classification). In contrast, deep learning-driven intelligent interpretation techniques (e.g., semantic segmentation) are insufficiently covered [11]. A survey of more than 20 universities in China that offer remote sensing courses found that fewer than 50% included practical teaching of deep learning for remote sensing image processing in their curricula. Consequently, students struggle to deal with complex ecological tasks, such as the fine classification of land use/land cover, which requires the integration of high-resolution imagery and deep learning, or the identification of small targets in optical/SAR images. On the other hand, the practical sessions often rely on closed platforms, such as ENVI/ArcGIS. Students can only complete processing through GUI clicks and cannot use open-source AI tools to solve real ecological problems. These issues hinder learning in remote sensing courses, including the detachment of theory from application, skills lagging behind technological advancements, and the inability to meet national ecological needs and global sustainable development challenges. These have become key bottlenecks restricting the cultivation of innovative talents.

In recent years, the deep integration of remote sensing technology and artificial intelligence has been profoundly changing the paradigm of earth science research. In particular, the development of deep learning has enabled the automatic interpretation of large-scale remote sensing data, environmental monitoring, and resource assessment [12,13]. However, approaching this issue solely from a technical perspective often overlooks its potential value in the field of education. To cultivate new-generation talent with both technical expertise and interdisciplinary perspectives, the seamless integration of artificial intelligence and remote sensing technology into higher education curricula has become a key issue in global higher education reform [14].

Internationally, the application of artificial intelligence in education is constantly expanding. From intelligent learning support systems [15] to data-driven curriculum design [16] and interdisciplinary project-based learning [17], technology is proving to be not only a research tool but also a catalyst for innovation in educational paradigms. In this process, the combination of remote sensing education and the Sustainable Development Goals (SDGs) is particularly noteworthy. For example, some universities abroad have integrated satellite remote sensing with topics such as ecological sustainability and climate change governance, exploring course practices based on real-world data [18]; in Australia, interdisciplinary courses emphasize enhancing students’ problem-solving and global competence through AI tools [19]. These practices provide valuable insights for the internationalization of remote sensing education. In contrast, China has made rapid progress in promoting “AI + Education” and New Engineering initiatives. However, existing research primarily focuses on the application of deep learning in remote sensing tasks [20]. This has led to limited systematic exploration of educational model innovation.

Based on the above, this study aims to explore how to organically integrate deep learning into remote sensing curricula under the backdrop of ecological civilization and sustainable development strategies, thereby driving higher education curriculum reform and interdisciplinary talent cultivation. This study first introduces commonly used deep learning models in remote sensing, then reconstructs the curriculum objective system by analyzing teaching pain points, and incorporates deep learning content. It proposes a five-dimensional linkage teaching method, and integrates various deep learning teaching tools and platforms to explore a trinity teaching reform path of “deep learning and remote sensing and ecological sustainability”. The effectiveness of the proposed curriculum teaching reform method will be validated through specific teaching cases and multi-dimensional assessment data. The purpose of this study is to bridge the gap between classroom theory and frontier applications, providing a reference for cultivating compound geographic information talents for sustainable development in the new era.

2. Classic Deep Learning Algorithms in Remote Sensing

In response to the urgent demands for complex features analysis, dynamic process assessment, and spatial correlation modeling in ecological environment monitoring, deep learning has emerged as a key technological driver revolutionizing the field of remote sensing. This advancement stems from its powerful capabilities in feature learning and nonlinear mapping. The following section introduces three deep learning models commonly applied in remote sensing image processing.

2.1. U-Net

U-Net is a convolutional neural network (CNN) architecture specifically designed for image segmentation [21]. It adopts a “U” shaped structure, which includes two main processes: encoding and decoding. In the network encoding stage, downsampling is performed using max pooling. This deepens the feature depth layer by layer. During decoding, the downsampled feature layers are merged with the upsampled feature layers to restore the original image resolution while preserving key feature information. The connection layer establishes skip connections between the encoder and decoder. It transmits high-resolution early features to the decoder to retain the original spatial information of the image. The network uses the ReLU function as the activation function of neurons, with the formula as follows:

$$ReLu\left(x\right) = \mathrm{m}\mathrm{a}\mathrm{x}(0, {\omega }^{{\rm T}}x + b)$$

(1)

where ω represents the weights that connect the input features to the neurons. x represents the input feature vector, and $b$ represents the bias used to adjust the output. This function outputs the input feature vector x from the upper-layer neural network to the lower-layer neural network. U-Net has performed well in segmentation and classification tasks for remote sensing images. Its unique skip connections can prevent information loss and enhance the recognition ability of small targets [22,23].

2.2. GCN

GCN was first proposed by Kipf et al. in 2017 [24]. It is a deep learning model that performs convolution operations on graph-structured data. It can effectively handle irregular data structures and is suitable for tasks such as node classification and graph classification [25]. When implementing GCN, the graph convolution operation aggregates node neighborhood information through the adjacency matrix A. Node features at each layer are generated from the features of nodes at the previous layer [26]. The normalized adjacency matrix is calculated as follows:

$$\widehat{A}=D^{-{\displaystyle \frac{1}{2}}}(A + I)D^{-{\displaystyle \frac{1}{2}}}$$

(2)

where $D$ is the degree matrix, the diagonal elements $D_{ii}= {\sum }_j{A}_{ij}$ represents the degree of node i, and $I$ is the unit matrix.

For the feature matrix $H^{\left(l\right)}$ of the nodes, the graph convolution operation aggregates features of the nodes using the adjacency matrix $\widehat{A}$. These aggregated features are then passed through the fully connected layer and the Softmax classifier to achieve pixel-level classification. The formula for the feature $H^{(l+1)}$ of the nodes in layer l + 1 is as follows:

$$H^{(l+1)} = \sigma \left(\widehat{A}{H}^{\left(l\right)}{W}^{\left(l\right)}\right)$$

(3)

Here, $H^{\left(l\right)}$ is the node feature matrix of the l-th layer, $W^{\left(l\right)}$ is the weight matrix of the l-th layer, $\sigma$ is the activation function, and ReLU is commonly used. In this formula, $\widehat{A}{H}^{\left(l\right)}$ aggregates the feature information of the node and its neighbors and performs a linear transformation through the weight matrix $W^{\left(l\right)}$, and, finally, nonlinearity is introduced through the nonlinear activation function $\sigma$.

GCN can effectively handle complex spatial relationships and utilize the local structure of graphs to improve classification or prediction accuracy. It has been widely applied in various fields, including image classification [27,28], semantic segmentation [29,30], change detection [31], etc.

2.3. Transformer

Transformer is an architecture based on self-attention mechanisms, initially used in natural language processing tasks [32]. It has significantly improved the ability to capture long-term dependencies through parallel processing mechanisms and self-attention mechanisms. The main body of the Transformer model consists of an encoder and a decoder, both of which adopt a multi-layer stacking design. The encoder relies on multi-head self-attention mechanisms and feedforward neural networks to perform deep encoding on input sequences and extract temporal global features.

The Transformer’s core operating mechanism originates from the self-attention principle. It dynamically integrates contextual information by calculating the correlation weights between each position and all other positions in the input sequence. From the perspective of matrix operation, its mathematical expression can be written as:

$$Attention\left(Q,K,V\right)=Softmax\left({\displaystyle \frac{Q{K}^T}{\sqrt{d_k}}}\right)V$$

(4)

where $Q$ is the query vector. It is used to search for relevant information in the key-value pairs and typically corresponds to the feature representation of the current image being processed. $K$ is the key vector. It stores the information available for querying and is used to calculate similarity with $Q$. $V$ is the value vector. It stores the specific content corresponding to the key vector and is used for weighted summation after obtaining the attention weights. $d_k$ is the dimension of the key vector $K$. $\sqrt{d_k}$ is introduced to scale the dot product result, and prevent the Softmax function from entering the gradient vanishing region when $d_k$ and the dot product results are too large.

The Transformer has a strong ability to deal with long-distance dependencies. It is suitable for large-scale datasets and offers higher parallel computing efficiency. They have gradually demonstrated their advantages in tasks such as remote sensing image analysis, object detection, and image segmentation [33,34].

3. Challenges in the Integration of Remote Sensing Education and Deep Learning

The current remote sensing curricula do not fully integrate deep learning methods, which is mainly reflected in the following aspects (as shown in Figure 1).

At the curriculum level, the current remote sensing curriculum suffers from significant structural rigidity and a disconnect from frontiers. Core courses, such as Principles of Remote Sensing, allocate approximately 75% of class hours to optical physics fundamentals, radiometric calibration, and traditional classification algorithms (e.g., maximum likelihood estimation and ISODATA clustering). In contrast, deep learning theories make up less than 10% of the content (according to our sampling statistics of remote sensing curriculum in Chinese universities in 2023). Interdisciplinary knowledge, such as machine learning theory and programming frameworks, is not systematically integrated into the curriculum. As a result, students struggle to build a comprehensive knowledge framework that connects remote sensing data, deep learning models, and ecological applications. This leads to two major issues: one is that the technical chain breaks. Machine learning fundamentals (e.g., gradient descent, backpropagation, etc.) are typically taught in computer science courses. They are not taught in conjunction with the characteristics of remote sensing data, such as spatial context dependence and multispectral channel correlations. Students struggle to understand how CNN convolutional kernels function in remote sensing data processing (e.g., extracting farmland features). They also find it difficult to grasp how Transformers analyze temporal changes in forest phenology. Another issue is the disconnection of ecological applications. For example, in the teaching of land classification, the focus remains on distinguishing coarse-grained targets such as farmland and construction land. However, these courses rarely cover frontier scenarios like precise carbon sink monitoring or ecological assessment within the context of sustainable development. The experimental results of the Remote Sensing Image Processing course in 2023 showed that the overall accuracy of farmland classification by students using traditional methods was less than 75%. This fails to meet the requirements for high-precision land cover monitoring and exposes the mismatch between the course content and complex ecological demands.

In terms of teaching methods and practices, the issue of tool dependency and disconnection from real scenarios is particularly prominent in practical teaching. Experimental classes are highly dependent on closed-source software such as ENVI/ERDAS. Students can only perform basic operations, such as K-Means clustering, maximum likelihood, and ISODATA classification, through GUI buttons. At the same time, real ecological projects require the use of PyTorch to build end-to-end intelligent processing pipelines. For example, in the wetland mapping project based on Sentinel-2 imagery, students first load Sentinel-2 data, customize a specific U-Net model, and ultimately output wetland mapping results. The survey reveals that fewer than one-fourth of the students can independently master the essential introductory deep learning skill of reading remote sensing data using the GDAL library and converting it into tensors (Tensor conversion) after completing the course. Additionally, teaching materials primarily use Landsat TM data (spatial resolution of 30 m). However, current ecological monitoring widely adopts sub-meter drone imagery and hyperspectral data. Students repeatedly practice basic scenarios such as urban expansion and land classification based on Landsat data. They have not yet been exposed to real tasks like time-series vegetation monitoring or forest disturbance detection. This results in a mismatch between their skills and industry demands.

At the teacher and resource level, there are contradictions among the knowledge structure of teachers, teaching materials, and the foundation of students. 80% of remote sensing teachers have graduated from traditional geographic/surveying majors and lack systematic AI training. When explaining the use of deep learning in specific case studies, only a few teachers can debug code on the spot. Most, however, rely on theoretical descriptions. This results in students being unable to grasp issues such as the Dropout layer for preventing overfitting and the Early Stopping strategy. Over 80% of mainstream textbooks lack case-based teaching related to deep learning. Moreover, their case data primarily consists of small, cropped images (256 × 256 pixels), which cannot support teaching large-scale ecological application cases. Students have varying levels of programming foundation. When faced with complex deep learning model development tasks, they often feel helpless and struggle to learn. Furthermore, the course assessments are still based on theoretical written exams, lacking assessment criteria for assessing students’ ability to apply deep learning models to solve specific cases.

The above analysis reveals that integrating deep learning organically into the remote sensing curriculum system is a significant challenge. It is also challenging to cultivate students’ innovative ability to solve ecological problems through intelligent means in the remote sensing curriculum teaching.

4. The Teaching Pathway for Integrating Deep Learning into Remote Sensing Curriculum

In response to the new demands of ecological civilization construction for remote sensing talents under the background of sustainable development, and addressing issues such as the disconnect between theory and practice and outdated tools in traditional teaching, this part has deeply studied the trinity teaching reform path of “deep learning and remote sensing and ecological sustainability” from four aspects: objectives, contents, methods, and tools, to enhance students’ ability to solve practical ecological problems.

4.1. Three-Stage Curriculum Objective System

The remote sensing curriculum reform integrating deep learning proposed in this study constructs a three-stage objective system of knowledge, ability, and literacy, forming a progressive talent cultivation path (as shown in Figure 2). The system is presented in a pyramid structure, with the knowledge objective that lays the theoretical foundation at the bottom, followed by the ability objective that emphasizes practical application, and the literacy objective that enhances thinking and values at the top. This progressive path aims to cultivate interdisciplinary talents who can support ecological civilization development.

The knowledge objective aims to bridge the gap between traditional remote sensing principles and cutting-edge artificial intelligence technologies. Students are required to master both the physical fundamentals of remote sensing and the mathematical core of deep learning (e.g., convolution operations, gradient backpropagation, etc.). They should also systematically understand the principles of classical and frontier models (e.g., U-Net and Transformer) and their adaptation mechanisms in typical ecological remote sensing tasks. For example, students should be able to explain from an algorithmic perspective why the local perception and weight-sharing characteristics of CNNs are suitable for vegetation fine classification of high-resolution imagery. They should also understand how the gated recurrent units of LSTMs effectively capture and utilize the long time-series dependencies in vegetation phenology.

The ability objective focuses on developing the ability to apply theoretical knowledge to eco-engineering practices that address real ecological challenges. Students must be proficient in using mainstream deep learning frameworks (PyTorch/TensorFlow) and geospatial processing libraries (GDAL). They should be able to independently complete the end-to-end problem-solving process, including data acquisition, preprocessing (e.g., multi-source data fusion, imbalanced sample handling, data enhancement, etc.), and model deployment. Further, students should be able to independently select and optimize model architectures, continuously improve the model’s generalization ability and recognition accuracy on complex terrains. This can be achieved through strategies like designing reasonable loss functions and adjusting learning rates. Ultimately, they should be able to write an ecological environment monitoring analysis report that combines technical depth and application value.

The literacy objective aims to go beyond specific technical operations. It focuses on cultivating students’ higher-order thinking and scientific spirit, and strives to develop students’ computational thinking and critical thinking. On the one hand, this approach guides students to decompose macro-level, complex ecological issues (such as regional biodiversity assessments) into data-driven computational tasks and to design algorithmic models to solve them. This process cultivates their ability to solve problems systematically. On the other hand, by thoroughly analyzing cases where models fail in real applications (such as misclassification of vegetation types due to mountain shadows or cloud interference), students are guided to dialectically examine the data dependency and generalization limitations of deep learning. The ultimate goal is to enable the students to establish correct scientific ethics and values through practical experience, and to form a sense of mission and high responsibility to apply cutting-edge intelligent technologies in serving the national ecological civilization construction.

The objective system designed for this curriculum reform is based on knowledge, centered on ability, and guided by literacy. It aims to guide students from mastering theory to applying it in practice, and then to elevating their thinking. This ultimately achieves a comprehensive improvement in their overall literacy.

4.2. Embedded Curriculum Content Fusion

Based on the existing content, an embedded integration strategy is adopted to establish a three-level modular system comprising foundational, advanced, and practical. Taking the course of Remote Sensing Image Processing as an example, Figure 3 shows the content modules before and after the curriculum reform. The pre-reform course content was divided into three parts: basic theory, image quality improvement, and image information extraction. The basic theory section included Chapter 1 (introduction) and Chapter 2 (mathematical foundations). The image quality improvement section included Chapter 3 (radiometric correction), Chapter 4 (geometric processing), and Chapter 5 (denoising and enhancement). The image information extraction section included Chapter 6 (segmentation and feature extraction), Chapter 7 (classification and change detection), and Chapter 8 (cartographic representation). Based on the original course content, the reformed incorporates an overview of the development history of deep learning in Chapter 1 and an introduction to typical deep learning principles and methods in Chapter 2, both of which constitute advanced theoretical foundations. In the practical phase, Chapter 6 incorporates the U-Net image segmentation case study to compare the advantages and disadvantages of deep learning methods versus traditional object-oriented methods in image segmentation. Chapter 7 adds a special topic on deep learning classification, allowing for a comparison of the accuracy differences between traditional SVM and deep learning in classification, and adds the Transformer model application module to demonstrate its advantages in capturing edge details through case studies in urban expansion monitoring.

A new 32 h specialized course on Ecological Remote Sensing Intelligent Analysis has been added. It includes the following content modules: (1) Foundations of Deep Learning, covering Python data structures and matrix operations; (2) Remote Sensing Data Preprocessing, such as multi-temporal data time series alignment and hyperspectral data dimensionality reduction; (3) Model Principles, focusing on the adaptability of CNN’s local receptive fields to the spatial correlation of remote sensing imagery; (4) Task Implementation, including code implementation for typical tasks such as classification and segmentation; (5) Comprehensive case studies. Each module includes three components: theoretical explanation, code demonstration, and self-practice. For example, in the U-Net land classification segmentation case, the matching logic between the encoder–decoder structure and land classification boundary features is first analyzed. This is followed by a demonstration of the code for loading the dataset using TorchGeo. Finally, students are asked to debug the IoU loss function parameters to optimize the segmentation results.

4.3. Five-Dimensional Linkage Teaching Model

This study developed a five-dimensional linkage teaching model (Figure 4). This model centers on project-based learning as its main line, running throughout the entire teaching process. It uses dual-track case-based studying as its core to foster students’ deep understanding. It employs modular training as a path to strengthen students’ engineering practical skills. It utilizes blended learning as a platform to optimize the allocation of learning time and space. And it takes collaborative learning as a bridge to cultivate students’ team spirit and sense of social responsibility. The five elements are interconnected and form an organic whole. This facilitates the deep integration of deep learning theory into the remote sensing curriculum system.

The curriculum is designed to address actual ecological issue chains, deeply integrating key concepts into project-based learning. Taking the Forestry Remote Sensing course as an example, this study designs a semester-long comprehensive project. This project is titled ‘Forest Disturbance Detection and Type Identification Using the GEE Platform and Deep Learning’. It spans 16 weeks and is divided into the four phases outlined in

Table 1. Project-based Learning Plan for Forestry Remote Sensing Course.

Time	Task	Main Content
1–4 weeks	Multi-source time series data processing	The GEE cloud platform is utilized to obtain long-term time series of Landsat and Sentinel-2 optical remote sensing data. Complete preprocessing such as cloud removal, mosaicking, cropping, radiation, and geometric correction.
5–8 weeks	Forest disturbance detection	Using LandTrendr, CCDC, and other algorithms in GEE, the pixel-level trend analysis of time series data is carried out. Automatically detect forest disturbance events, including key parameters such as the time of occurrence, duration, and variation amplitude.
9–12 weeks	Forest disturbance type identification	The U-Net model is applied to semantic segmentation to classify pixels in disturbed areas, and accurate identification of different disturbance types, such as fire, logging, diseases, and insect pests, is realized. Encourage the introduction of unsupervised contrastive learning (such as SimCLR) for feature pre-training, and explore methods to improve the robustness and generalization capabilities of models when labeled data is limited.
13–16 weeks	Results compilation and accuracy verification	Integrate disturbance detection results with type identification results, and perform accuracy verification. Generate high-precision distribution maps of forest disturbance and disturbance types, then finish the project reports.

. The project aims to guide students in comprehensively applying cloud platform computing power, time-series remote sensing analysis, and deep learning models to solve the core challenges in the dynamic monitoring of forest ecosystems. Furthermore, it can cultivate students’ rigorous scientific research literacy and high ecological responsibility.

To deepen students’ understanding of model selection, we have established a dual-track case library centered on the core framework of traditional machine learning vs. deep learning. The traditional cases are drawn from highly cited papers, such as the application of SVM and random forests in wetland classification. Students dissect their feature extraction and parameter optimization logic. The deep learning cases focus on teachers’ latest research achievements in ecological remote sensing, for instance, U-Net-based precise extraction of Spartina alterniflora and Transformer-based spatio-temporal evolution analysis of Poyang Lake. This helps guide students to analyze end-to-end learning paradigms and advanced parameter adjustment strategies, including dynamic learning rates and hybrid loss functions. Through comparisons of the same scenario using different methods, such as SVM and U-Net in coastal wetland classification tasks, or traditional Wishart classifiers and GCN graph models in polarimetric SAR classification tasks, students are guided to select and evaluate technical approaches scientifically and dialectically, from data characteristics, task objectives, and model generalization capabilities.

Aiming at the differences in students’ programming foundations, a three-layer progressive code template is designed to enable modular training tailored to individual skill levels. The basic version provides fully commented, complete code that students can run directly to gain an intuitive understanding of the model workflow quickly. The advanced version reserves key functions or model structure interfaces, requiring students to modify hyperparameters, such as the convolution kernel size or learning rate, or replace modules, to experience the model optimization process. The challenge version only provides the dataset and a clear ecological monitoring task, encouraging students to independently review literature, select, and build an appropriate deep learning model. For example, in the forest fire detection training, the basic version allows students to directly run the ResNet model, the advanced version requires modifying the convolution kernel size, and the challenge version requires students to select the model architecture independently. This layered design aims to lower the entry barrier and provide innovative exploration opportunities for capable students. It also effectively enhances the engineering coding and model implementation capabilities of all students.

The “2 + 3” blended learning model combines online independent learning with offline inquiry discussions. Before class, students watch short videos provided by teachers on the online learning platform. These videos cover topics such as model principles and key code explanations. This allows students to complete self-directed learning of foundational knowledge and submit pre-class thinking questions (2 h). Class time is transformed into high-value project-based inquiry workshops. Here, students debug code, discuss problems, and advance projects in groups. Teachers act as facilitators and guides, providing on-site guidance to focus on solving common and individual issues students encounter in practical scenarios, such as data loading, model non-convergence, and overfitting. This facilitates deep knowledge construction (3 h).

The project adopts a collaborative learning model featuring heterogeneous grouping and role-based assignments. Each group (4–5 people) includes roles such as data scientists (responsible for data preprocessing and analysis), algorithm engineers (responsible for model construction and debugging), and ecological analysts (responsible for result interpretation and report writing). Students select roles based on their interests and strengths. The entire project process is managed through GitHub for version control and task collaboration. Commit records, issue discussions, and other elements serve as essential references for process-based assessment. This mode helps develop students’ communication and collaboration skills as well as project management capabilities.

4.4. Deep Learning Teaching Tools

To ensure the smooth implementation of teaching reforms, we designed and constructed an integrated teaching support system for remote sensing courses incorporating deep learning. This system combines software, platform, and data, as shown in Figure 5.

To accommodate the needs of different learning stages, the software tools are configured hierarchically. The basic tools use Python (v3.9/v3.10), GDAL (v3.x), and Jupyter Notebook (v6.5.x). Python and GDAL are used for data format conversion, band operations, and other basic processing. Jupyter Notebook serves as an interactive environment to facilitate real-time code execution and visualization. This lowers the operational threshold for students. The core framework primarily uses PyTorch (v1.13.1/v2.0.1), with TensorFlow (v2.10.x) as a supplement. PyTorch’s dynamic graph mechanism is easier for students to understand and debug. This makes it suitable for introductory teaching and capable of covering 90% of model training needs. The specialized libraries include TorchGeo (v0.5.x) and Rasterio (v1.3.x). TorchGeo enables efficient loading of remote sensing data. Rasterio simplifies operations such as band reading and writing. This reduces repetitive coding work for students in data processing.

For teaching and research tasks of different complexity, the computing power platform adopts the dual-track system. Basic training relies on local CPUs and Baidu AI Studio. Baidu AI Studio provides free and stable GPUs, meeting the needs of basic experiments such as 512 × 512 pixel-level image classification. Complex projects are deployed on the school’s self-built GPU cluster and Alibaba Cloud PAI-DSW, which obtains V100 graphics card support through the unified application of educational resource packages from the school laboratory. The ecological remote sensing dataset management system adopts a combination of Label Studio and Git. Label Studio is responsible for data annotation, while Git implements version control. The system also integrates an automatic generation function for index files, allowing direct output of index files compliant with framework standards.

Data serves as the bridge connecting theory and practice. The constructed database encompasses three-level datasets: (1) The basic database primarily includes open datasets such as Landsat-8/9 and Sentinel-1/2 series imageries; (2) The thematic database consists of the Global Wetland Dataset, Forest Fire Dataset, etc.; (3) The self-made database mainly focuses on regional ecological issues. It consists of ecological data collaboratively labeled by teachers and students, with no more than 500 samples per category. This data is obtained through a “teacher pre-labeling and student correction” model. Examples include local forest disturbance imagery and landslide data from southwestern mountainous regions. All data is accompanied by simplified metadata documents detailing data sources, processing steps, and typical application scenarios.

In this teaching tool support system, software, platforms, and databases cooperate to serve teaching practices. Software tools provide technical support for data processing and model training. Computing platforms ensure the resources needed for algorithm operation. Databases provide the material foundation for the entire teaching activity. The three closely cooperate to assist students in learning from basic operations to complex project practices.

5. Practical Outcomes and Effectiveness Assessment

To validate the effectiveness of the remote sensing curriculum reform based on deep learning in the context of ecological civilization construction, this study conducted teaching practices in remote sensing courses for surveying and mapping majors at our university. This involved a control group in 2024 and an experimental group in 2025. The experimental group consists of 29 students who followed the reform plan designed by this study. In comparison, the control group comprises 30 students who followed the traditional teaching model, which primarily involves teacher-led lectures and traditional verification experiments. All participants voluntarily participated in the study and provided informed consent. Data collection and analysis strictly adhered to confidentiality and privacy protection principles. This section will demonstrate the effectiveness of the teaching reform systematically from three aspects: teaching practice case demonstrations, multi-dimensional assessment system construction, and implementation effectiveness analysis.

5.1. Hierarchical Case Practice Design

Students in this major take required and elective courses in remote sensing during their undergraduate studies, including Principles of Remote Sensing, Forestry Remote Sensing, Ecological Remote Sensing Intelligent Analysis, Remote Sensing Image Processing, and Microwave Remote Sensing. The practical components in the curriculum focused closely on key remote sensing application scenarios in ecological civilization construction, incorporating a series of tiered, progressively challenging comprehensive case studies. These case studies aim to guide students in applying frontier deep learning methods to real ecological and environmental issues, thereby enhancing their overall abilities. The following are some case study outcomes from the experimental group.

(1)

Dynamic monitoring of Spartina alterniflora invasion and removal in Jiangsu coastal area based on U-Net model.

This case study serves as an introductory and core application of semantic segmentation technology, aiming to guide students in applying cutting-edge artificial intelligence methods to address the critical ecological issue of biological invasion control. Through a comprehensive practical cycle, students will systematically master semantic segmentation technology and its capabilities for precise mapping and dynamic analysis in ecological management. Specific practical requirements are as follows: Students must obtain multi-source remote sensing imagery covering different years from the Jiangsu coastal region, including Landsat 8 (optical), Sentinel-1 (radar), and Sentinel-2 (optical), and perform standardized data preprocessing. Based on the phenological characteristics of Spartina alterniflora, guide students to analyze and determine the optimal extraction time window; simultaneously, utilize machine learning methods such as Random Forest for feature selection to enhance the model’s discriminative ability for Spartina alterniflora. Students are required to apply deep learning semantic segmentation models, such as U-Net, for automated identification and high-precision mapping of Spartina alterniflora. After completing the distribution extraction of Spartina alterniflora in 2022, further analysis will be conducted on the removal dynamics and effectiveness of Spartina alterniflora by integrating the removal operation information from seven ports in the region in 2023, to assess the progress of invasive species control. Figure 6 shows the technical process implemented by the students, and Figure 7 displays the results of their practical work. The results show that by integrating multi-source remote sensing data and the U-Net model, the overall extraction accuracy of Spartina alterniflora reached 94%. This high accuracy not only validates the efficiency of the selected deep learning methods in complex wetland environments but also fully demonstrates the students’ ability to apply their knowledge and tools to solve real-world ecological and environmental problems. This case study successfully integrates theoretical instruction with cutting-edge technological practice and ecological management needs, effectively enhancing students’ deep learning application capabilities, complex remote sensing data processing capabilities, and practical skills in addressing specific ecological and environmental issues.

(2)

Forest disturbance type identification based on SimCLR unsupervised contrastive learning and U-Net model.

This case study aims to address the challenge of obtaining high-quality labeled data in forest ecosystem monitoring and guide students to explore the application of unsupervised representation learning techniques in the field of remote sensing. The core process and student learning path are as follows: Students first collect and process long-term Landsat remote sensing image data, use the LandTrendr algorithm to conduct preliminary disturbance detection of forests, identify areas that have changed, and provide a spatial scope for subsequent classification. To address the challenge of lacking labeled data, students must employ SimCLR for feature pre-training. Through unsupervised contrastive learning, the model can automatically learn robust and discriminative forest feature representations from massive amounts of unlabeled time-series image data, significantly reducing reliance on manually labeled data. Combining the pre-trained effective features, students further utilize semantic segmentation models such as U-Net to achieve classification and mapping of forest disturbance types (e.g., logging, fires, landslides) across three distinct regions. Figure 8 is the schematic diagram of the U-Net framework. The students’ results (Figure 9) show that through this unsupervised learning workflow, the students successfully classified complex forest disturbances with high accuracy, achieving an overall classification accuracy of over 80%. This achievement is particularly noteworthy because it was achieved under the practical constraint of scarce labeled data, fully demonstrating the potential of unsupervised learning methods in addressing data scarcity in the field of ecological and environmental science. This conclusion is consistent with that in Reference [23]. This case study successfully cultivated students’ ability to apply cutting-edge unsupervised learning techniques for advanced feature extraction, model building, and innovative problem-solving in complex, data-constrained scenarios, significantly enhancing their research and practical skills.

(3)

Dynamic monitoring of Poyang Lake water area based on Transformer model.

This case study addresses the challenge of processing large amounts of time-series remote sensing data to accurately capture hydrological dynamics in the assessment of lake wetland ecological health. The experiment requires students to explore Transformer and its variant models in depth, leveraging their advantages in handling long time series dependencies and complex spatial contextual information to achieve intelligent monitoring and prediction of large lake hydrological dynamics. The core teaching components and skill development pathways of the case study are as follows: Students must acquire and preprocess long time-series Sentinel-2 remote sensing imagery data from the Poyang Lake region. Attempt to construct or improve Transformer-based variant models (such as Swin-UNet, etc.) to adapt to the special requirements of lake water body extraction. Utilize the constructed models to achieve precise extraction and temporal change analysis of water body areas in Poyang Lake across different months, revealing the seasonal and interannual fluctuation patterns of lake water levels and water body ranges. The students’ results show that by applying Swin-UNet (Figure 10 displays the network structure of Swin-Unet), the students successfully obtained water body distribution information for Poyang Lake from January to December, with an overall extraction accuracy of over 95%, as displayed in Figure 11. This result not only demonstrates the powerful potential of Transformer models in complex hydrological dynamic monitoring but also fully showcases the students’ capabilities in handling large-scale, complex spatiotemporal remote sensing. This is consistent with the findings in Reference [33]. Through this case study, the students gained a deep understanding of how to apply cutting-edge deep learning technologies, particularly the Transformer architecture, to address challenging hydrological geography problems, laying a solid foundation of skills and innovative thinking for future participation in global water resource management and ecological wetland conservation.

(4)

Polarimetric SAR image classification based on GCN networks.

This case study is designed to address the complexity of polarimetric SAR image processing in the Microwave Remote Sensing course, guiding students to explore how to use deep learning methods to improve interpretation accuracy. The core teaching components and competency development pathways of this case study are as follows: Students will gain a deep understanding of the physical scattering mechanisms and complexity of polarimetric SAR data and learn how to perform data preprocessing. Guide students in constructing and applying GCN models to fully utilize the spatial neighborhood information and topological relationships between pixels in the image. Students will utilize L-band AIRSAR fully polarimetric SAR data, with farmland in the Flevoland region as the study area, to achieve high-precision classification of different crop types and compare performance with traditional classification algorithms. The students’ case study results (Figure 12) clearly demonstrate the superiority of GCN in processing complex polarimetric SAR data. Compared to the traditional Wishart classification algorithm, the GCN-based method achieved a 33.38% improvement in overall accuracy and a 0.37 increase in the Kappa coefficient in the classification of Flevoland farmland. This case study not only highlights GCN’s powerful capabilities in processing complex SAR data but also directly validates the immense potential of deep learning in enhancing the accuracy of microwave remote sensing interpretation. Through this, students have expanded their capabilities in processing new types of remote sensing data and deepened their understanding of cutting-edge GNN algorithms.

5.2. Construction of a Multi-Dimensional Comprehensive Assessment System

To scientifically and comprehensively evaluate the effectiveness of this teaching reform, this study moves beyond the limitations of traditional single academic performance assessments. Instead, it establishes an innovative, comprehensive assessment system centered on student development, integrating multiple dimensions such as knowledge, ability, and literacy. This system integrates multi-source data generated by students during the course learning process, including unit tests, programming logs, experiment reports, and concept maps. This data objectively reflects students’ growth in five areas: learning ability, application ability, practical ability, ecological thinking, and comprehensive problem-solving ability.

Table 2. Multi-dimensional assessment framework.

Dimension	Learning Outcome	Assessment Indicator	Data Source	Weight
Knowledge dimension	Students can comprehend and articulate the fundamental principles and core concepts of deep learning models.	Model learning foundation: degree of student’s cognition and depth of understanding of core deep learning model principles and key concepts.	Unit tests	0.1
			Concept map analysis	0.1
	Students can master the application methods of deep learning models in remote sensing image processing and accurately complete typical tasks.	Remote sensing application capability: in typical tasks, the model’s accuracy, generalization capability, and efficiency.	Experiment report	0.18
			Model evaluation	0.12
Ability dimension	Students can write standardized, efficient, readable, and engineering-compliant deep learning remote sensing code.	Engineering practice ability: standardization, readability, modularity, and performance of student-written code.	Automatic code checking tool	0.2
	Students can independently analyze, identify, and efficiently debug problems encountered in deep learning models in remote sensing applications.	Problem-solving ability: efficiency of model debugging (e.g., time taken to find and solve problems), thoroughness of problem resolution, and innovation of methods.	Jupyter Notebook logs	0.2
Literacy dimension	Students can demonstrate in-depth data analysis capabilities, critical thinking, and ecological civilization awareness in ecological remote sensing projects.	Ecological thinking: in project reports, the depth of students’ ecological analysis of multi-source remote sensing data, logical rigor, and breadth of understanding of ecological issues.	PPT presentation	0.1

presents five indicators across the three dimensions of this system, along with their respective weights. These weights were determined through discussions and surveys by several full-time teachers who have relevant teaching experience in remote sensing.

Based on the above table, the achievement value of each indicator can be calculated using the following formula,

$$M={\displaystyle \frac{\frac{1}{n}{\sum }_{i=1}^n{\sum }_{j=1}^{1,2}{M}_{ij}\times {\omega }_j}{M_{max}}}$$

(5)

where $n$ is the total number of students in the class, $M_{ij}$ is the score of a particular data source $j$ ($j$ = 1 or 2) in the indicator, ${\omega }_j$ represents the weight of the data source, and $M_{max}$ represents the maximum score of that indicator.

The above quantitative indicators can dynamically present the achievement values of individual and class goals in the teaching process of this course. They also provide teachers with real-time, detailed teaching results. This helps teachers identify weaknesses in teaching, adjust teaching strategies promptly, and continuously optimize teaching and provide personalized tutoring.

5.3. Analysis of the Effectiveness of Teaching Reform Implementation

Based on the multi-dimensional assessment system and the indicator achievement calculation method proposed in the previous section, we can calculate the achievement of each indicator in the remote sensing curriculum system teaching. The results show that this teaching reform has achieved remarkable results in cultivating composite remote sensing talents that meet the needs of ecological civilization construction in the context of sustainable development.

As can be seen from Figure 13, the experimental group students effectively achieved the above five indicators (>0.6). Compared with the control group, students in the experimental group achieved significant improvements in their mastery of model learning fundamentals and remote sensing application skills, with increases of 6% and 9%, respectively (Figure 14). In specific projects, students demonstrated a substantial enhancement in their technical implementation capabilities. For example, in the project monitoring the invasion and removal dynamics of Spartina alterniflora, the U-Net model developed by experimental group students achieved an average extraction accuracy of 94%, surpassing the control group’s results. In the polarimetric SAR image classification task, the average classification accuracy was improved by 30%. Additionally, through assessment of students’ project code and model principle understanding, experimental group students demonstrated superior performance in terms of model understanding depth and efficiency, with code standardization improving from 68% to 85%, and project completion efficiency increasing by about 28% (Figure 14).

Students in the experimental group have constructed more complex and deeper knowledge networks by the end of the courses, establishing stronger connections between remote sensing, deep learning, and ecological issues. They demonstrated greater transferability and innovation when dealing with complex problems. For example, in the case of forest disturbance type identification, two groups successfully combined unsupervised contrastive learning SimCLR with U-Net for few-shot learning, thus effectively solving the problem of scarce labeled data. In the Poyang Lake water area dynamic monitoring case study, a group attempted to combine Transformer with LSTM to construct a water level-area response model. They achieved an average absolute percentage error (MAPE) of less than 10%. These innovative attempts indicated that students have evolved from simple model users to researchers who can propose, validate, and solve complex problems.

Based on an analysis of the ecological interpretation section in the PPT presentations, students in the experimental group generally incorporated profound insights and ethical considerations into their reports. These insights covered ecological issues such as the ecological hazards of Spartina alterniflora, forest disturbance and restoration, and the carbon sink value of forests. Several students stated that through their active participation in these projects and case studies, they “truly understand the irreplaceable role of remote sensing technology in the construction of green mountains and clear waters”. They also clarified their professional mission to dedicate themselves to ecological conservation in the future. This fully demonstrates that ecology-oriented teaching design has successfully achieved the organic integration of knowledge transmission, skill development, and value cultivation.

The practical outcomes of experimental class students extend beyond the classroom. Some outstanding works have been converted into practical applications or scientific research outputs. For example, the polarimetric SAR image classification method based on GCN networks has been applied for an invention patent. Additionally, forest disturbance type identification based on U-Net has been incorporated into the regional ecological resource monitoring system. Two groups of students drafted academic papers based on their case or project outcomes and submitted them to relevant scholarly journals. These outcomes not only validate the effectiveness of teaching but also demonstrate the actual contributions of the courses to ecological civilization construction.

5.4. Limitations and Suggested Mitigation Strategies

The current educational reform research has yielded satisfactory results, but several limitations remain. First, the proposed reform model has only proven effective in surveying and mapping undergraduate education at Chinese universities. Its universality and applicability in institutions with different resource conditions or cultural characteristics, or in other courses, still need further verification. This is especially true for schools with insufficient computing resources, where implementation may be limited. Second, although the model has strong technical support, it presents a steep learning curve for students with limited programming backgrounds, potentially causing some students to struggle to keep up with the pace, thereby affecting their learning outcomes. Additionally, the teaching process may overly emphasize the cultivation of technical skills, potentially leading to insufficient attention to conceptual understanding and affecting the development of students’ comprehensive abilities. Therefore, to address these shortcomings, the following improvement measures are recommended: First, design hierarchical learning support for beginners, providing resources tailored to students with different learning backgrounds, such as implementing teaching content in stages and gradually increasing technical difficulty. Secondly, conduct systematic teacher training programs to enhance their ability to guide students in adapting to new teaching models and ensure teaching quality. Thirdly, balance conceptual and technical instruction to ensure that students not only master technical skills but also deeply understand the underlying theories and concepts, thereby improving overall teaching effectiveness. Additionally, it is recommended that future research be expanded to broader educational contexts and cultural backgrounds for validation, and that long-term follow-up studies be conducted to comprehensively assess the long-term effects of the reforms and their adaptability under different conditions, thereby providing a more robust theoretical and practical foundation for the continuous optimization of teaching reforms.

6. Conclusions

This study focuses on the teaching reform of remote sensing courses in the background of ecological civilization construction. We constructed a teaching model and a multi-dimensional assessment system by deeply integrating deep learning technology into teaching practice. These aim to cultivate composite remote sensing talents who meet the needs of sustainable development. This reform scheme was driven by solving real ecological problems and sought to enhance students’ ability to solve actual ecological problems with intelligent technology. It has important theoretical and practical significance for bridging the gap between frontier knowledge and classroom teaching and promoting the innovative development of remote sensing education. The research conclusions are as follows:

(1)

This study addressed the disconnect between traditional remote sensing teaching content, cutting-edge technology, and ecological needs. It proposed a teaching reform scheme driven by solving actual ecological problems and implemented through project-based practice. This effectively restructured the teaching content modules, closely combining them with ecological civilization construction and Sustainable Development Goals. This led to a more logical and practical knowledge system.

(2)

A hierarchical project-based learning model was designed. Specific cases, such as “Forest disturbance detection and type identification based on GEE and deep learning”, run through the teaching process. Students are guided through the entire workflow. This includes multi-source time-series data processing, LandTrendr disturbance detection, and U-Net disturbance type identification. They also explore cutting-edge technologies (e.g., SimCLR). This approach enhances students’ practical ability to use intelligent technologies to solve complex ecological problems.

(3)

A diversified teaching method system of five-dimensional linkage, such as case teaching, modular training, and hybrid teaching, was constructed. Combined with mainstream tools such as PyTorch/TensorFlow, TorchGeo, and cloud platforms, the system provided an immersive, efficient, and convenient learning environment for students. This effectively enhanced their engineering practice skills and independent learning capabilities.

(4)

A multi-dimensional comprehensive assessment system integrating the three dimensions of knowledge, ability, and literacy was established. Core indicators were refined, and multiple data sources were used for objective assessment. A quantitative M-value calculation formula was proposed. This enabled the precise evaluation and feedback on each indicator’s achievement for individual students and classes. This effectively supported the continuous improvement of teaching quality and verified the effectiveness of sustainable talent cultivation.

The trinity teaching reform concept and method of “deep learning and remote sensing and ecological sustainability” proposed in this research provide a reference framework and practical approach for the reform of remote sensing courses. The findings of this study provide valuable insights for education policymakers and university administrators in developing policies to cultivate interdisciplinary talent capable of serving the Sustainable Development Goals. However, in the practical implementation of the teaching reform methods proposed in this paper, it is important to note the limitations mentioned in the text, such as how institutions with limited resources can implement these methods, the steep learning curve faced by students with weak programming foundations, and the risk of overemphasizing technical skills at the expense of conceptual understanding. The paper provides specific strategies to address these challenges. This study also acknowledges that as deep learning technologies and ecological issues continue to evolve, future curriculum content and teaching methods must be continuously updated and iterated to adapt to future complex challenges.