Land Use and Land Cover Maps for Stream Water Quality Assessment in Spatial Buffers: A Systematic Review of Recent Trends (2020–2024)

Abstract

Assessing the impact of land use and land cover (LULC) on water quality (WQ) is central to land-based environmental research. Adhering to the Preferred Reporting Items for Systematic Reviews and Meta-Analyses guidelines, this study analyses recent trends using LULC maps to assess stream WQ within buffers, focusing on papers published between 2020 and 2024. It identifies relevant remote sensing practices for LULC mapping, landscape metrics, WQ physicochemical parameters, statistical techniques for correlating LULC and WQ, and conventions for configuring buffers. Materials include Scopus, Web of Science, and Atlas.ti, which support both qualitative data analysis and Conversational Artificial Intelligence (CAI) tasks via its integration with OpenAI’s large language models. The methodology highlights creating a bibliographic database, coding, CAI, and validating prompts. Official maps and visual or digital interpretations of optical imagery provided inputs for LULC. Classifiers from earlier generations have shaped LULC cartography. The most employed WQ parameters were phosphorus, total nitrogen, and pH. The three most referenced landscape metrics were the Largest Patch Index, Patch Density, and Landscape Shape Index. The literature mainly relied on Redundancy Analysis, Principal Component Analysis, and alternative correlation approaches. Buffer configurations varied in size. CAI facilitated an agile systematic review; however, it encountered challenges related to a phenomenon known as hallucination, which hampers its optimal performance.

1. Introduction

The recent breakthroughs in acquiring, processing, and distributing spatial data are transforming the landscape of land-based environmental research. By integrating various data sources, researchers can better address specific contextual questions, such as the impact of land use and land cover (LULC) on water quality (WQ). The emergence of significant technological advancements coincides with the growing volume of literature indexed in Scopus and Web of Science (WOS), featuring terms such as “Land Use” or “Land Cover” alongside “Water Quality” in their titles, abstracts, or keywords from 2020 to 2024. Figure 1 illustrates 4408 articles in Scopus and 9698 in WOS, which may exhibit some overlap. Both databases show a general growth trend, with WOS consistently reporting higher counts.

Numerous studies have primarily focused on assessing the impact of LULC on the physical, chemical, and biological indicators of WQ. This research has been conducted across different water bodies, including oceans, rivers, reservoirs, and groundwater. Initiatives [1,2,3,4] have demonstrated that different types of LULC exhibit unique effects on WQ. This evaluation considers geographical context, recognising that parameters can differ significantly between rural and urban areas or landscape compositions. It is essential to choose relevant time frames—whether days, months, seasons, or even years—and define suitable scales of analysis, employing spatial units such as catchments, sub-catchments, or buffers.

The effectiveness of buffers, referred to in this systematic review as either spatial units for analysis or normative buffer zones (BZSs), in mitigating the impacts of LULC on WQ is well documented in the literature [5,6,7,8]. Buffers serve as protective areas that filter pollutants, reduce runoff, and ultimately enhance WQ by diminishing the adverse effects of LULC on nearby water bodies. Considering the issues, a literature review is suitable for synthesising the significant scientific evidence and elucidating existing research patterns, trends, and gaps. Several examples with distinctive yet complementary focuses emerge prominently.

Cheng et al. [9] undertook a comprehensive review of remote sensing (RS) and statistical methodologies for establishing the relationships between LULC and WQ. This analysis predominantly encompassed the literature published from the late 1990s to the 2010s. The principal findings indicate that large-scale, coarse-spatial-resolution data, particularly obtained from the Landsat programme, had emerged as a significant resource in global research due to their capability for long-term continuous observation, medium temporal resolution, and extensive global coverage. The quantitative relationship between LULC and WQ was evaluated through various mathematical and statistical techniques, including correlation analysis, regression analysis, redundancy analysis, and mixed models.

Wu et al. [10] enlightened the effects of riparian BZSs on WQ and agricultural non-point source pollution control (NPSP) and how their structures and mechanisms contribute to controlling water pollution. The study concluded that riparian BZSs are practical for water safety, NPSP control, and creating a suitable environment for terrestrial and aquatic species. Moreover, they could reduce water temperature due to the shading effect and sustain water habitats by acting as a climate adaptation measure. However, the long-term influence of the riparian BZSs on trapping NPSP pollutants and soil properties is a research gap.

Locke [11] comprehensively reviewed statistical methods for modelling the interactions between LULC and WQ. The study highlights key findings on statistical approaches, WQ indicators, LULC data, and the interplay between temporal and spatial perspectives. Significant gaps were identified, including persistent uncertainty about the influence of spatial scale and landscape configuration and the potentially confounding effects of additional variables that, if not adequately addressed, can compromise the accuracy of statistical models. Moreover, a geographical bias in the existing literature underscores the need for further research in ecologically and climatically diverse regions, particularly in underrepresented areas of the Global South.

Shi et al. [12] analysed 625 regional studies across 63 countries to investigate how LULC changes affect global WQ. The study found that urban expansion negatively impacted WQ, while increased forest cover had a positive effect, particularly in low-latitude forest regions. Wetlands also show a positive correlation with organic matter levels in water. The study highlighted consistent and varying effects of these changes on WQ, noting an intensified trend since the 1990s, which is vital for addressing global water scarcity. Statistical methods revealed that urban expansion significantly affected WQ based on factors such as latitude, seasonality, and climate zones.

These reviews underscored the importance of spatial scale and landscape configuration in understanding how different LULC types impact WQ. Buffers were widely recognised as effective in reducing agricultural non-point source pollution. However, there are significant gaps in the long-term evaluation of their effectiveness. These findings highlighted the need for integrated land management strategies that consider the long-term effects of LULC on WQ, especially in less-studied areas. Substantial scientific documentation on these interactions could be effectively processed and analysed using tools such as qualitative data analysis software (QDAS) and natural language processing (NLP).

QDAS aids researchers in integrating various document sources, identifying patterns, and extracting insights that contribute to a comprehensive understanding of the literature [13]. Furthermore, it enhances the transparency and reproducibility of research findings [14].

NLP, a specialised area within Artificial Intelligence (AI), enables computers to understand and communicate in human language. This technology can effectively identify key themes by extracting relevant textual elements [15] and providing contextually relevant results [16]. NLP is a significant component of state-of-the-art tools such as Conversational Intelligence (CAI), which refers to technologies like chatbots that enable user dialogue. These systems utilise extensive datasets and advanced language models. They also employ NLP capabilities to simulate human interactions, recognise speech and text inputs, and interpret meanings across various languages.

This investigation systematically reviewed recent research articles (2020–2024) that explore the effects of LULC on stream WQ, with a particular focus on analyses conducted within cited buffers. By employing analytical tools such as QDAS and CAI, this study examined key topics following the research outcomes (ROs) outlined below:

RO1. The technical foundations of RS used to acquire or generate LULC data. This includes the utilisation of pre-existing maps, satellite imagery, classification algorithms, and accuracy assessment statistics.

RO2. The ten most frequently applied landscape metrics (LMTSSs).

RO3. The ten most used physicochemical WQ parameters.

RO4. The most recurrently employed statistical procedures (STATs) for assessing correlations or predictive relationships between LULC and WQ variables.

RO5. The prevailing practices for defining the geometric configuration of buffers, particularly regarding their shape and size.

RO1 and RO2 were accomplished through the implementation of QDAS. RO3, RO4, and RO5 were tackled using CAI. Allocating a research segment to QDAS facilitated a more traditional yet agile and flexible methodology. The integration of CAI for the remaining components enhanced agility and fostered innovation in the management and interpretation of extensive textual data. This research employed a CAI approach as a method of generative AI, with the results assessed through a validation process described in Section 3.5.

In today’s rapidly evolving technological landscape, this study addressed a gap by providing an updated synthesis of the essential spatial data and methodologies used to analyse the effects of LULC on WQ. The review examined the primary data sources and RS practices employed in the construction of LULC databases, as well as the variables most frequently used to evaluate WQ across environments comprising natural (e.g., forests, wetlands) and anthropogenic (e.g., urban, agricultural) features. The emphasis is placed on buffers, which constitute a critical spatial dimension from ecological, geographical, and management perspectives. Through this lens, the analysis revealed the patterns and trends that advance the understanding of the spatial nexus between LULC and WQ.

In contrast to previous reviews, this study adopted a systematic approach that integrated selected evidence to identify the principal factors shaping these interactions. This analysis highlights practices and identifies challenges and opportunities for enhancing methodological transparency.

1.1. Limitations and Potential Bias

Systematic reviews employ a transparent and reproducible methodology for analysing existing evidence. They follow specific inclusion and exclusion criteria. However, these criteria may cause some relevant issues to be overlooked.

This review omitted studies involving spatial dimensions such as groundwater, marine environments, or lentic water bodies. It also did not consider biological indicators and contaminants beyond physicochemical parameters, such as microplastics. This means the analysis did not infer paramount understanding, such as ecological responses derived from biological indicators. Metrics derived from macroinvertebrates, algae, or fish are essential for assessing ecosystem health and provide insights that physicochemical data alone cannot capture.

It also disregarded initiatives considering simulations or hydrological modelling, which represent significant innovations that could accurately represent local hydrological and WQ scenarios under validated conditions. This study did not analyse the effects of other environmental variables that influence WQ, such as geological formation, soil composition, and climate conditions. These factors can interact with LULC in complex ways, shaping hydrological processes, nutrient dynamics, and pollutant transport. Their omission implies that the patterns identified may only provide a partial view of the drivers of WQ. Further details regarding these exclusions, as well as additional exclusions and their justifications, can be found in Section 3.1.2.

This review did not provide an independent assessment of the methodological rigour or data quality of the studies included, as it aimed primarily to synthesise patterns, approaches, and findings across the literature. Consequently, the review relied on the quality standards assumed in the peer review process of the included publications, recognising that this approach may introduce limitations concerning the robustness and comparability of the reported evidence.

Constraining this systematic review to peer-reviewed papers published in English created a bias by omitting insights from non-peer-reviewed sources and significant research published in other languages. As a result, this may exclude critical regional evidence, particularly from countries in the Global South, where pertinent scientific outcomes may exist but do not meet the linguistic criteria for inclusion. Focusing on the 2020 to 2024 period may overlook earlier effective practices that remain relevant.

While this investigation employed CAI, a thorough examination of the underlying algorithms, including neural architectures and training methodologies, is beyond the scope of this review. This study did not seek to train or develop specialised NLP models; instead, it utilised OpenAI’s existing capabilities through Atlas.ti 25.0.1 (ATI) to augment various findings and analyses.

1.2. Key Terms for the Systematic Review

Defining key terms is crucial for an efficacious systematic review. It creates a clear framework and improves understanding for both researchers and readers. Establishing definitions for these terms ensures consistency throughout the review process, which aids in accurately interpreting findings. This research focused on five key terms (

Table 1. Key terms, definitions, and sources.

Key Terms	Definitions	Sources
Land use /land cover	Land cover is a “Physical and biological cover of the Earth’s surface including artificial surfaces, agricultural areas, forests, (semi-)natural areas, wetlands, water bodies…” while land use comprises the “…Territory characterised according to its current and future planned functional dimension or socio-economic purpose (e.g., residential, industrial, commercial, agricultural, forestry, recreational”.	[17]
	“Land cover is defined as the observed physical cover on the Earth’s surface, including vegetation (natural or planted) and human constructions. Land use is characterised by the arrangements, activities and inputs people undertake in a certain land cover type to produce, change, or maintain it. Definition of land use in this way establishes a direct link between land cover and the actions of people in their environment”.	[18]
Stream	Water that flows naturally along a fixed route formed by a channel cut into rock or ground.	[19]
Water quality	WQ is a complex concept that includes water’s physical, chemical, and biological properties, varying depending on its intended use. Establishing universal standards for WQ is challenging, as water safe for human consumption may not meet the requirements for agricultural purposes and vice versa.	[20]
Buffer (generic)	The term “buffer” is frequently employed within the field of geographic information technologies. In this context, it refers to all areas that are located within a specified distance from a river or its segments. Applied research investigating the effects of LULC on WQ may utilise either linear or circular buffers.	The authors
Buffer zones	The concept generally takes on a more functional and administrative role compared with the generic definition of a buffer. In this sense, BZSs are “… designated areas used to protect sensitive landscape patches (e.g., wetlands, wildlife reserves) from negative external pressures … should be based on the desired ecological functions, landscape context, and external pressures”.	[21]

): LULC, stream, WQ, buffer (generic spatial domain), and BZSs.

2. Materials

2.1. Repositories: Scopus and Web of Science

Using more than one repository for systematic reviews improves coverage [22], which may help address regional disparities in journal access [23].

Scopus is an extensive abstract and citation database that spans multiple disciplines. Its key advantages include a meticulously curated collection of abstracts and citations. Its strong focus on emerging markets provides access to high-quality research from developing countries that may not be available in other databases. The platform is source-neutral and curated by a global panel of independent subject matter experts [24].

WOS is a comprehensive repository that provides access to multiple citation databases, indexing various scholarly studies across various disciplines. It enables researchers to analyse trends and discover influential academic work. WOS emphasises rigorous quality control and offers a variety of tools for citation analysis and research evaluation [25].

Google Scholar is a widespread search engine for finding research; however, it is unsuitable for this research. It has limited subject indexing, and the documents are not always peer reviewed. Also, its web interface does not allow users to search and offers little control or functionality.

Ultimately, Scopus and WOS serve as key repositories for querying the scientific literature. They offer an interface that facilitates the filtering of documentation based on various criteria, including, but not limited to, key terms, publication date, and research area.

2.2. Atlas.ti

ATI is a QDAS that is helpful in systematic reviews. It offers tools for organising and analysing findings from multiple studies, enabling researchers to code and categorise data efficiently. The software leverages OpenAI’s generative pre-trained transformer (GPT) models to power its CAI, which interprets and contextualises user queries. When using CAI features, users submit their documents and questions to OpenAI’s servers for processing and responses.

3. Methods

This research adhered to the most recent version of the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) guidelines [26]. Moreover, further studies [27,28,29] provided supplementary insights and contextual depth that elucidated the methodological foundations and epistemological relevance of systematic reviews. These guidelines facilitated transparent reporting on the review’s rationale, methods, and findings. A protocol, pending registration, was followed to integrate the guidelines and workflows; this ensured consistent and rigorous outcomes.

The methods comprised six distinct phases: the creation of the bibliographic database, parameter selection and justification, coding, prompt drafting and execution, the validation of prompt outcomes, and the conversion of unstructured text into structured text.

3.1. Creation of the Bibliographic Database

In line with [26], Figure 2 illustrates a simplified process for building the bibliographic database, which included three main stages: identification, screening, and database consolidation. Each stage is further explained below for clarity.

3.1.1. Identification

On 30 January 2025, an exploratory search examined the Scopus and WOS repositories. Various search terms and Boolean operators, detailed in

Table 2. Repository, search scope, terms, and Boolean operators in the search engine.

Repository	Search Scope	Terms and Boolean Operators
Scopus	Article title, abstract, and keywords.	TITLE-ABS-KEY ((“Land use” OR “land cover”) AND “water quality” AND “buffer”) AND PUBYEAR > 2019 AND PUBYEAR < 2025 AND (LIMIT-TO(DOCTYPE, “ar”)) AND (LIMIT-TO (LANGUAGE, “English”))
Web of Science	Title, abstract, and indexing.	“Land use” OR “Land cover” (Topic) and “water quality” AND “buffer” (Topic)

, guided the search within each platform’s engine. The repositories differ slightly in search functionalities: Scopus allows language filtering directly within the Boolean query, while WOS provides this option through a separate filter. The search targeted articles published in English between 2020 and 2024 that included the terms “Land Use”, “Land Cover”, “Water Quality”, and “Buffer” in their titles and abstracts.

Although terms such as “maps”, “mapping”, “cartography”, “spatial technologies”, or “stream” might have refined the search, these were intentionally excluded. Their addition often resulted in overlooking significant findings.

Scopus retrieved 207 articles, while WOS provided 340. After eliminating duplicates, 378 unique documents were identified.

3.1.2. Screening

Although using AI to assist in screening documents for systematic reviews is technically possible, there is still no explicit agreement on its effectiveness. For instance, one study [30] demonstrated that AI-assisted screening achieves a sensitivity and specificity comparable to manual screening. Another investigation [31] reported that automated methods attained a high sensitivity. Similarly, Ref. [32] noted that AI can reduce workload and expedite the title and abstract screening stage. However, several limitations must be acknowledged. Further steps are necessary to accurately identify trial registrations or studies that meet the inclusion criteria for systematic reviews [33]. Human oversight may be crucial for enhancing accuracy and preventing errors [34], which underscores the importance of a cautious and critical approach to the application of existing AI systems [35].

To avoid potential constraints, the authors used a thorough human screening approach. This process included setting exclusion criteria and manually selecting relevant articles from both institutional and open-access sources.

Table 3. Exclusion criteria and rationale.

Criteria	Rationale
1. Do not examine the influences of LULC on WQ or those that do not utilise LULC maps	This investigation focuses explicitly on the effects of LULC on WQ; using corresponding maps is mandatory.
2. Fail to consider buffers or BZSs	Buffers and BZSs are key spatial dimensions in this research.
3. Reviews, meta-analyses, and theoretical articles	This exclusion prevents duplication of information, as reviews summarise findings from primary studies already included. Theoretical articles lack the specific evidence needed to address the research questions.
4. Studies exclusively focused on urban dynamics	This research aims to analyse documentation of urban and rural environments, making it appropriate for deriving insights from holistic settings rather than solely focusing on urban areas.
5. Studies employing simulation or hypothetical scenarios	These studies rely on informatics and statistical approaches whose comprehensive analysis goes beyond the aims and focus of this review.
6. Investigations evaluating biological indicators	The research focuses on physicochemical parameters.
7. Studies considering marine, lentic, or groundwater bodies	Marine, lentic, or groundwater bodies are systems with distinct hydrological and biogeochemical processes whose behaviour and dynamics extend beyond the focus of this research.

outlines the rationale for the exclusion criteria. The criteria included (1) sources that did not examine the influences of LULC on WQ or that did not utilise LULC maps; (2) studies that did not take buffers into account; (3) review articles, meta-analyses, and theoretical papers; (4) research that exclusively analysed urban dynamics; (5) studies that relied on simulation methods or hypothetical scenarios; (6) investigations that assessed biological indicators derived from biological communities or pollutants beyond physical or chemical parameters; (7) projects that considered marine, lentic, or groundwater bodies. These criteria ensure that the selected research aligns with this study’s objectives.

The process of manually screening and selecting accessible articles from both institutional and open sources proceeded in two distinct phases. In the first phase, 103 manuscripts were excluded based on various criteria assessed through their titles and abstracts. In the second phase, 275 articles were identified as potential candidates.

The authors evaluated the documents for eligibility and agreed on the final selection of references to ensure accuracy and alignment with the review’s objectives.

Supplementary File S1 presents several examples of studies [A, B, C, D, E, F, G, H] that appeared to meet the inclusion criteria but were ultimately excluded following a thorough analysis.

Ultimately, 70 manuscripts were deemed suitable for this systematic review.

3.1.3. Bibliographic Database Consolidation

The 70 articles were downloaded as PDFs from the journal’s website and categorised by source and publication year. They were then stored in a local folder for further processing in ATI.

3.2. Parameter Selection and Justification

The parameter selection and criteria are consistent with the study’s ROs. RO1 addressed the identification of LULC maps and mapping strategies, including imagery, classification algorithms, and accuracy assessments. RO2 and RO3 concentrated on determining the ten most applied LMTSs and WQ variables in buffer-scale analyses, offering a concise overview of prevailing practices. RO4 incorporated STATs due to their widespread application in evaluating the impacts of LULC on WQ. Finally, RO5 explored buffer configurations to emphasise the spatial dimension of LULC–WQ relationships.

The parameters were extracted based on the coding or prompt drafting for CAI execution, as detailed below.

3.3. Coding

The process involved assigning meaningful labels, known as “codes,” to different sections of text that describe the technical foundations for RO1 and the 10 most used LMTSs for RO2. For example, to identify elements such as pre-existing maps, satellite imagery, classification algorithms, and accuracy assessment statistics used in the documentation, corresponding elements were labelled once for every study employing this approach. A similar process was used to identify LMTSs.

A key challenge during coding was standardising terminology for concepts that were essentially the same but inconsistently named in the literature. For example, “Interspersion Juxtaposition Index” and “Interspersion and Juxtaposition Index” were consolidated under a single term. Similarly, “Largest Plaque Index” and “Largest Patch Index” were unified, with the latter chosen in both cases to ensure clarity and consistency.

After completing the coding process, an .xlsx file was produced for further analysis.

3.4. Prompt Drafting and Conversational Artificial Intelligence Execution

A prompt serves as a user input that guides a CAI in generating responses. Within the CAI context, prompts are essential because they allow the AI to capture the nuances of human language, including context and intent.

After considering different options,

Table 4. Prompts designed to support various research objectives.

ROs	Prompt ID	Prompt’s Text
RO3	PR3	List the physicochemical water quality parameters that were examined in relation to LULC.
RO4	PR4	Identify only the statistical procedures explicitly employed to assess correlations or predictive relationships between land use and water quality variables in the reviewed document. Focus strictly on correlation analyses (e.g., Pearson, Spearman), regression models (e.g., linear, multiple regression), and constrained ordination techniques such as Redundancy Analysis (RDA) when applied to these relationships. Exclude any statistical tests comparing group means (e.g., ANOVA, t-tests, Kruskal–Wallis), descriptive statistics, normality assessments, or inferential methods unrelated to correlation, regression, or constrained ordination. Provide a direct quotation from the text to support your response, along with the corresponding citation.
RO5	PR5	Identify the buffer distances and geometric configurations (circular or linear) used in this research. Additionally, I would like you to determine which buffer distance has the most influential impact on water quality. If possible, include a direct quote from the text to support your answer, along with the appropriate citation.

illustrates the selected prompts that addressed RO3, RO4, and RO5. These were tested on 10% of the documentation included in the systematic review, specifically refs. [3,4,36,37,38,39,40].

The responses provided valuable Supplementary Information, addressed errors effectively, ensured relevance, and delivered meaningful answers [41]. In contrast, several previously tested prompts encountered issues such as inappropriate responses, factual inaccuracies, and hallucinations, which present significant barriers in the realm of CAI [42].

Using the CAI, the selected prompts were applied to all documents in the systematic review.

3.5. Validation of Prompt Outcomes

To check the reliability of the prompt outcomes, a validation was performed on 30% of the sample. This included 21 randomly selected studies [7,43,44,45,46,47,48,49,50,51,52,53,54,55,56,57,58,59,60,61,62]. The results generated by the CAI were compared with those obtained through two different validation methods, each focusing on either identification or interpretation.

The first mechanism evaluated the percentage of correct identifications (POI) made by the CAI using parameters from RO3 and statistics from RO4. Following the CAI’s identification of targets for both RO3 and RO4, a manual verification process was conducted to check the accuracy of these identifications across all samples. This process aggregated correct identifications and errors, classifying errors as commissions (incorrectly including unlisted elements) or omissions (overlooking existing elements). This approach facilitated the calculation of the POI.

The second mechanism involved a framework for evaluating the CAI interpretation of RO5. This framework employed a scoring system outlined in

Table 5. Criterion and scoring.

Criterion	Score If Met	Score If Unmet
Directness/coherence	2	0
Completeness	1	0
Clarify	1	0
Proper quotation	1	0

, which assessed directness, coherence, completeness, clarity, and proper quotation usage. Points were assigned as follows: two points for directness and coherence, and one point each for the other factors.

After evaluating all samples, the scores were averaged and compared to reference values to determine the suitability of the CAI interpretations according to the following ranges:

• Interpretations with an average score of 0 to 3 were discarded.
• Scores between 3.1 and 4.5 indicated that interpretations need substantial refinement but could serve as a guiding framework.
• Scores from 4.6 to 5 suggested that interpretations were adequate and could serve as a guide with minor improvements.

3.6. Conversion of Unstructured Text to Structured Text

This step focused on the initial results from the CAI responses. Unstructured text was converted into a structured format by organising the raw data into a table, enabling easier analysis and information retrieval. This process involved filtering out irrelevant words, categorising content, and standardising the text to British English for consistency in the research.

4. Results and Discussions

This systematic review included 31 journals listed in Supplementary Table S2. Figure 3 summarised their contributions, showing the number of papers published by each journal along with their specific SJR areas and Q rankings as indicated in the figure’s legend.

Water leads with 14.29% of publications, followed by Ecological Indicators (11.43%) and Science of the Total Environment (10.00%). Journal of Cleaner Production accounts for 8.57%, and Environmental Science and Pollution Research represents 7.14%. Journal of Environmental Management holds 5.71%, while journals such as Catena, Environmental Pollution, Water Research, Journal of Hydrology, and Polish Journal of Environmental Studies each contribute 2.86%. The “Others” category comprises 28.57%, reflecting the diversity of research topics.

The journals reviewed span a broad range of disciplines, underscoring the interdisciplinary relevance of the topic in environmental science and water management. By 2024, many of these journals had achieved Q1 rankings in the Scimago Journal and Country Rank, demonstrating their significant impact within the academic community. This reflects their quality, as the selected papers are essential for producing important research outcomes.

The analysis showed a notable concentration of case studies in particular countries. As depicted in Figure 4, China leads with a 61.48% share, which translates to a contribution of 43 papers. This finding aligns with the comprehensive review by [63], which highlights China’s predominant participation in relevant peer-reviewed LULC papers, and the conclusions by [64] that underscore the country’s pre-eminence in global natural sciences research. Following China, Iran accounts for 5.71% (four papers), while Poland and the United States each provide 4.28% (three papers). South Africa, in turn, holds a 2.78% share, equivalent to two papers. The countries/regions that each contributed one paper (1.43%) include Argentina, Bangladesh, Burkina Faso, Costa Rica, Ethiopia, Germany, Ghana, Mexico, Nigeria, Pakistan, Portugal, Serbia, Taiwan, Turkey, and Vietnam.

The restricted number of studies focusing on specific geographical contexts, particularly in the Global South, highlighted the findings from references [10,11]. The predominance of studies from China (over 60% of the total) revealed the strong interest in LULC–WQ indexed research in that country. It could also be a result of the extensive availability of technological resources, such as comprehensive monitoring networks, the creation of LULC datasets, and large-scale data infrastructures. Although this offers valuable evidence, it also constrains the generalisability of the findings. Buffer configurations and metric sensitivity may differ across regions depending on land management practices, policy frameworks, and levels of technological development, which can hinder their transferability to other contexts.

Section 4.1 and Section 4.2 examine each study included in the systematic review, clarifying whether the research relies on existing LULC maps or develops new ones using diverse interpretation schemes.

4.1. Land Use/Land Cover Pre-Existing Maps

Thirty-eight studies utilised pre-existing maps produced by government agencies or academic institutions. The data were primarily derived from the Landsat legacy, which includes Landsat 4/5 Thematic Mapper (TM), Landsat 7 Enhanced Thematic Mapper (ETM+), and Landsat 8 Operational Land Imager (OLI), as well as resources from the Sentinel-2 Multispectral Instrument (S2MSI).

Table 6. Reference, year of publication, and the maps’ source.

Year of Publication	Reference	Maps’ Source
2020	[44]	Land use vegetation data sourced from the National Institute of Statistics and Geography in Mexico
	[45]	Coordination of Information on the Environment Land Cover (CLC)
	[65]	United States Department of Agriculture Cropland Data Layers
	[66]	GIS-oriented information
2021	[3]	Environmental and Ecological Science Data Centre for West China, National Natural Science Foundation of China
	[46]	Highly detailed pan-European land cover map from 2015, underpinned by the LULC Area Frame Survey
	[67]	Geo Data Wisconsin databases for Dane and Columbia Counties
	[68]	National Land Cover Database—United States Geological Survey
	[69]	Ningbo Academy of Agricultural Sciences
	[70]
2022	[37]	Data from the Statistical Yearbook of Jiangxi Province and the China Statistical Yearbook
	[49]	GlobeLand30
	[51]	CLC and the S2 Land Cover Map of Europe
	[53]	Finer Resolution Observation and Monitoring Land Cover dataset
	[71]	Database of Topographic Objects
	[72]	Geomatics Information Centre of Hunan Province
	[73]	30 m annual China Land Cover Dataset
2023	[54]	The CLC
	[55]	30 m annual Land Cover Dataset and its associated dynamics in China
	[56]	Resource and Environmental Science and Data Centre of the Chinese Academy of Sciences (RESDC)
	[57]	GlobeLand30
	[58]
	[74]	RESDC
	[75]	Coordination of Information on the Environment Land Cover (CLC)
	[76]	30 m annual China Land Cover Dataset (CLCD)
2024	[39]	ESA World Cover, the Argentinian National Crop Map, and the LULC Map for the Córdoba City Peri-urban Area
	[40]	National land use data sourced from Earth System Science Data
	[62]	Information from the Natural Resources Bureaus of Zhenxiong and Weixin counties
	[77]	ESRI’s Sentinel global land cover data
	[78]	Cross-resolution Land Cover data
	[36]	South African National Land Cover
	[79]
	[80]	LULC data from Tsinghua University
	[81]	Unnamed land use maps
	[82]	GlobeLand30 (GL30)
	[83]
	[84]	RESDC
	[85]

presents references organised by the year of publication. It also includes the sources of maps, providing an overview of the temporal distribution and data inputs across the studies.

The studies revealed a diverse range of LULC data sources across regions, with a notable reliance on national and global datasets such as CLC, GL30, CLCD, and RESDC. National agencies and research institutes play a central role in producing and providing access to LULC information tailored to regional contexts. While pre-existing maps produced by reputable organisations are regarded as reliable, the analysis suggests that the reviewed literature generally does not provide explicit reporting of their accuracy metrics.

4.2. Recent Approaches for Land Use/Land Cover Mapping for Water Quality Assessment

4.2.1. Imagery for Land Use/Land Cover Classification

Thirty-two studies created maps, primarily using Landsat and S2MSI. The specific details are presented below:

• OLI was identified in twelve studies [4,43,47,48,52,59,86,87,88,89,90,91].
• S2MSI was helpful in three studies [61,92,93].
• Google Earth services [94,95] and the Gaofen-2 High-Resolution Optical Satellite [60,96] were used in two studies each.
• Aerial photographs from the Ethiopian Mapping Authority [97], the Rapid Eye satellite [98], and the China–Brazil Earth Resources Satellite [99] were used once each.
• A total of ten investigations employed more than one source to ensure thorough coverage of longer time frames that exceed the scope of a single mission. Specifically, the studies utilised TM and OLI [38,50,100,101], TM and ETM+ [102,103], ETM+ and OLI [7,104], TM and S2MSI [105], and a complementary usage of TM, ETM+, OLI, and S2MSI [106].

These findings highlighted the enduring legacy of the Landsat programme and the significance of Copernicus S2 as a critical resource for providing free access to data with unprecedented temporal and medium spatial resolution. This point is consistent with recent systematic reviews highlighting a firm reliance on optical sensors [107,108]. Details on the specific bands used and the calculation methods for indices from satellite imagery are rarely available.

While this research incorporated the latest scientific evidence, it is important to note that multisource RS is a prominent approach that includes alternatives to optical sources [109] and has yet to be thoroughly examined. The absence of active systems, such as synthetic aperture radar (SAR) or Light Detection and Ranging (LiDAR), is particularly notable.

4.2.2. Procedures for Land Use/Land Cover Delineation and Accuracy Metrics

Consistent with [110], the studies commonly classified LULC into agricultural land, urban/built-up areas, commercial forestry, grasslands, natural forests, and wetlands. Additionally, mining activities have received considerable attention.

Twenty-five studies outlined methods for LULC delineation. Figure 5 illustrates the three primary techniques used: supervised classification (15 instances), unsupervised classification (3 cases), and visual interpretation (7 instances). Additional information regarding the specific algorithms is provided below.

• Supervised classification was the most frequently reported, with fifteen studies employing it. Within this group, four sources [52,87,105,106] generally referred to supervised classification. One study [39] mentioned the Decision Tree (DT), while [101] focused on a Support Vector Machine (SVM). Three sources [47,92,104] considered Random Forests (RF) and six investigations [4,38,43,59,86,93] applied Maximum Likelihood (ML).
• Three sources reported using unsupervised approaches. Ref. [102] mentioned it as “unsupervised classification”, while two sources [7,90] expressly mentioned using self-organising data analysis clustering (SOC).
• Seven studies [60,61,69,70,89,95,97] highlighted visual interpretation.

Most procedures in this research used older algorithms like ML or DT, similar to the studies reviewed in [11] that focused on earlier years. Few studies examined recent traditional classifiers such as RF or SVM, while visual interpretation was a simple and effective solution.

There was a notable absence of reporting on contemporary deep learning techniques, including convolutional neural networks, deep neural networks, fully convolutional neural networks, recurrent neural networks, or autoencoders [111,112].

Nineteen studies [4,7,36,38,39,43,47,51,57,59,61,76,79,89,100,101,102,104,105] reported one or more metrics such as the overall accuracy (OA), Kappa index (Kappa), and user’s and producer’s accuracy. The documented accuracy metrics suggest optimal performance. Frequently, the lower OA ranged between 75% and 82%, while the higher values extended up to 94.50%. Regarding the Kappa values, they ranged from 0.69 to 0.87, with a more prevalent range noted between 0.82 and 0.86. In general, most studies that classified imagery provided accuracy statistics. However, in a few instances, this information was omitted.

Approximately 27% of the studies (19 out of 70) reported using desktop software for classification. Among these, solutions developed by the Environmental Systems Research Institute (ESRI), particularly ArcMap versions 10.1, 10.2, 10.8, and ArcGIS Pro, were the most frequently adopted, representing approximately 16% of the total cases [58,60,61,69,70,81,89,90,97,103,105]. Additionally, ArcMap 10.2 was combined with the Earth Resources Data Analysis System (ERDAS, version 9.2) [86], while ERDAS version 14 was employed in other studies [59,105,106]. The Environment for Visualising Images (ENVI), specifically versions 5.1 and 5.3, was used in about 6% of the studies [4,52,87,102]. This evidences the scientific community’s trust in the cited software.

None of the reviewed studies reported using cloud-based platforms such as Google Earth Engine or Microsoft Planetary Computer for classification. It evidences a lack of integration of recently adopted RS innovations into other specialised research areas. This disparity is likely due to disciplinary silos, where advancements in RS tend to remain confined to specific communities, thereby hindering their dissemination into applied domains. Additional contributing factors may include differing research priorities, a greater emphasis on domain-specific innovations rather than methodological advancements, and limited access to the computational resources or training required to implement cutting-edge RS techniques. Collectively, these constraints may impede the broader application and innovative potential of state-of-the-art RS approaches across various disciplines.

4.3. The Ten Most Employed Landscape Metrics

Out of the 70 studies reviewed, 27 (38.57%) used LMTSs. Figure 6 presents the ten most frequently used metrics in decreasing order: Largest Patch Index (LPI), Patch Density (PD), Landscape Shape Index (LSI), Shannon Diversity Index (SHDI), Interspersion and Juxtaposition Index (IJI), Contagion Index (CONTAG), Edge Density (ED), Aggregation Index (AI), Percentage of Landscape (PLAND), Number of Patches (NP), Patch Cohesion Index (COHESION), and Mean Euclidean Nearest Neighbour Distance (ENN_MN). It should be noted that CONTAG and ED are tied for sixth place, while NP and COHESION share ninth place.

These findings support those in ref. [113], although that study did not focus on ranking. The metrics identified in this review include the LPI, PD, LSI, SHDI, CONTAG, and ED. They have demonstrated consistent relevance over the years

The LPI is the most frequently occurring metric in the analysis, with an incidence of 27. The research shows that it effectively reflects the intensity of human activities [100]. It is widely recognised as a suitable measure for correlating LULC and WQ across various scales [92] and for clarifying seasonal variations in WQ [56]. Additionally, the LPI has contributed to understanding the multi-scale effects of landscape patterns on nitrogen (N) and phosphorus (P) dynamics [102].

The second most reported metric is PD, mentioned 24 times. This indicator of landscape fragmentation significantly influences WQ by increasing surface runoff and transporting pollutants in fragmented land use patterns, such as agricultural or urban settings, while disrupting natural filtration systems like riparian zones [114]. PD helps clarify how specific LULC types affect WQ parameters [101]. For example, the density of forest and grassland patches in riparian zones negatively correlates with total nitrogen (TN) levels, indicating that denser vegetated patches intercept nitrogen before it reaches waterways [58]. Additionally, PD is valuable for assessing LULC-WQ relationships across different spatial scales, including watersheds and buffers [70], as well as temporal scales, especially seasonal variations [102].

LSI, referenced 19 times, ranks as the third most utilised metric. This metric serves as a standardised measure of total edge or Edge Density, adjusted for landscape size [4]. The LSI has been pivotal in evaluating how LULC influences WQ across different seasons [92] and in explaining notable spatial discrepancies [82]. Additionally, the LSI provides reference values for specific LULC types, such as urban areas, and illustrates their adverse effects on declining WQ [56].

SHDI ranks fourth, with 17 scholarly articles discussing it. According to [115], SHDI reflects landscape diversity and heterogeneity through two main aspects: the types of patches and their distribution within the landscape. Empirical studies reveal a negative correlation between SHDI and WQ pollution indicators, suggesting that higher SHDI values, which signify greater landscape diversity, are associated with improved WQ in rivers. For instance, increased diversity has been shown to help disperse heavy metal contamination [49]. Additionally, SHDI has a strong positive correlation with sediment quality indices, highlighting the significant adverse effects of agricultural and urban LULC on WQ [52].

The IJI appears 16 times, making it the fifth most frequently analysed metric. The IJI aids in proposing landscape threshold intervals for stream risk management (SRM) [95] and identifies the influence of specific LULC types on certain WQ attributes [92,93].

CONTAG and ED share the sixth place in the ranking. CONTAG is effectively correlated with pond water N and P concentrations [60] and heavy metals [49]. It highlights significant variations in cation concentrations, showing that landscape fragmentation tends to increase cation export, regardless of specific LULC types [59]. ED is crucial for understanding how different LULC types, especially natural forests, impact WQ in agricultural catchments [116]. It has also been used to carry out successful threshold analyses of landscape patterns and WQ at optimal spatial scales [101]. A notable study [93] found that higher forest edge densities correlate with improved WQ, suggesting that complex woodlands along river corridors and gullies enhance the landscape’s water filtration capacity.

AI, the seventh most frequently used metric, facilitates the understanding of variations in WQ at different scales [93] and identifies the influence of specific crops, such as rice paddies [74].

PLAND refers to the proportion of a specific patch type within the total landscape area. It is an essential metric for understanding landscape composition, as it indicates the extent to which a particular land cover or habitat type occupies the overall area. PLAND represents the proportion of each LULC class within the landscape or area under analysis. For instance, ref. [103] utilised it to explain how LULC influences WQ.

Both the NP and the COHESION are positioned ninth. Research [96] indicates that increased NP correlates with poorer WQ and greater landscape fragmentation. In contrast, high forest cohesion levels are linked to reduced contamination [101].

The ENN_MN ranks tenth. Studies show that in urban areas, higher ENN_MN levels are associated with increased electrical conductivity [93]. This metric is also helpful in assessing the impacts of LULC at the watershed level [58].

The significance of metrics such as LPI, PD, and LSI in the literature highlights the crucial role of spatial configuration in mediating the effects of LULC on WQ. At the LULC class level, these metrics offer vital insights into structural dominance (LPI), fragmentation (PD), and edge complexity (LSI), all of which influence hydrological connectivity, surface runoff, and pollutant transport.

Metrics like SHDI and IJI reflect landscape heterogeneity and spatial arrangement, linking compositional diversity with ecological resilience and buffering capacity against contaminants.

The frequent use of CONTAG and ED underlines the importance of landscape aggregation and edge effects, especially in detecting threshold responses in agricultural and urban catchments. Additionally, AI, PLAND, and COHESION operate as valuable indicators of landscape connectivity and the dominance of functional land covers—like forests or rice paddies—that enhance nutrient retention and water filtration. Finally, ENN_MN illustrates patch spatial isolation, often linked to urban sprawl and degraded WQ.

The LMTSs effectively assess the impact of various environmental policies across nations and targets specific scientific objectives related to WQ monitoring.

4.4. Prompts, Responses, and Accuracy

4.4.1. Identification of Water Quality Parameters

In the analysis of RO3, the CAI identified WQ variables on 184 occasions from the samples, with varying distributions across different studies. One study examined two variables, while another analysed twenty-five.

Table 7. Example of prompt and response in RO3 analysis.

Prompt

Response

PR2

Chat Answer: The study analysed several physicochemical parameters of water quality to connect these parameters to Land Use/Land Cover (LULC) in the study area. Here is a summary of the findings: Analysed Physicochemical Parameters of Water Quality Dissolved Oxygen (DO)—measured in mg/L Biochemical Oxygen Demand (BOD5)—measured in mg/L pH Electrical Conductivity (EC)—measured in µS/cm Temperature—measured in °C Total Dissolved Solids (TDS)—measured in mg/L Total Suspended Solids (TSS)—measured in mg/L Total Solids (TS)—measured in mg/L.

Note: for clarity, the table comprises solely the responses provided by the CAI, leaving out any supplementary text from the interface.

provides an example of a CAI based on [105].

Out of 184 occasions, 182 were correctly identified, achieving a nearly 99% accuracy rate. There was only one omission and one commission. This high level of accuracy confirms that the WQ variables identified by the CAI are reliable. Instances of misleading responses were rare and did not significantly impact the quality of the automated outcomes. Given the CAI’s optimal performance, the automated responses are suitable for interpretation and topic extraction.

4.4.2. Identification of Recurrently Employed Statistical Procedures for Assessing Correlations or Predictive Relationships Between Land Use/Land Cover and Water Quality Variables

In the analysis of RO4, similar to RO3, only the selected samples were evaluated to assess the effectiveness of the CAI. The CAI identified STATs that connected LULC influences on WQ 59 times, with varied methods across studies. One study used a single STAT, while others used up to four.

Table 8. Example of prompt and response in RO4 analysis.

Prompt	Source	Response
PR4	[78]	Chat Answer: The statistical procedures explicitly employed to assess correlations or predictive relationships between land use and water quality variables in the reviewed document include: 1. Redundancy Analysis (RDA): “The redundancy analysis (RDA) was conducted to determine the influence of integrated landscape patterns on seasonal water quality.” “The RDA quantified the effects of comprehensive landscape patterns on water quality.” 2. Partial Redundancy Analysis (pRDA): “The total explained the amount of RDA analysis includes the independent and interactive contributions of different landscape categories, which can be evaluated using partial RDA (pRDA).” “The partial redundancy analysis (pRDA) shows the explanation of landscape composition, landscape configuration, and physiographic metrics to water quality at different spatial scales and seasons.” These methods focus on understanding the relationships between landscape metrics and water quality, specifically through the lens of redundancy analysis techniques.
	[69]	Chat Answer: The statistical procedures explicitly employed to assess correlations or predictive relationships between land use and water quality variables in the reviewed document include: 1. Redundancy Analysis (RDA): “The DCA results showed that the longest gradient length of the four ordination axes was <3. Thus, the redundancy analysis (RDA) was adopted to determine the positive/negative relationship between landscape metrics and water quality parameters”. 2. Partial Redundancy Analysis (pRDA): “Partial RDA (pRDA) was further applied to clarify and determine the relative importance and pure effects of the physiography, landscape composition, and landscape configuration on water quality”. 3. Nonparametric Change-Point Analysis (nCPA): “The nonparametric change-point analysis (nCPA) is very effective in evaluating the positions of abrupt change points relevant to the water quality parameters and landscape metrics”. These methods focus on understanding the relationships between landscape metrics and water quality, specifically through correlation and regression techniques.

Note: for clarity, the table comprises solely the responses provided by the CAI, leaving out any supplementary text from the interface.

, referencing [69,78], presents an optimal response where STATs are correctly identified and a misleading response where nonparametric change-point analysis is wrongly classified as a STAT.

The test in Section 3.4, which covered 10% of the documentation, yielded positive results. However, the validation in Section 3.5 reached a POI of approximately 69.59%. This suggests unpredictability in the CAI’s performance on this specific matter. It generated multiple commissions, around 30.41%, by including irrelevant statistics. For instance, it referenced procedures such as “Retention Effect Calculation”, “Dependent Sample t-test”, and “Partial Mantel Tests”, which do not align with the requirements. The POI demonstrates that summing the frequency of each measure based on CAI responses is inappropriate. Nonetheless, these analyses enabled the manual identification of the STATs and their practical applications for extracting key topics.

The challenges highlighted in RO4 pertain to hallucinations and the model’s insufficient domain-specific training. Hallucinations occur when generative AI produces factually incorrect, nonsensical, or unrelated results [117]. This typically occurs when the model attempts to fill knowledge gaps by fabricating information rather than acknowledging its limitations [41,42]. This variability likely stems from the model’s limited specialisation in specific domains. OpenAI cautions against this risk, noting that the lack of a definitive source of truth during training can lead to responses that seem plausible but are ultimately inaccurate or misleading [118]. Supporting these findings, other studies [119,120,121] have reported varying accuracy across different subject-specific queries. However, the adaptive nature of generative AI implies that ongoing user interaction could enhance its performance.

4.4.3. Analysis of Buffer Configurations

The RO5 performance analysis exemplifies two circumstances that highlight the accuracy of the CAI and instances where errors occurred. This evaluation aims to elucidate the strengths and weaknesses of the approach.

Table 9. Examples of prompts and responses in RO5 analysis.

Prompt	Source	Response
R5	[49]	Chat Answer: In the research, the following buffer distances and geometric configurations were identified: Buffer Distances and Geometric Configurations 1. Buffer Distances: Point buffers at distances of 100 m, 300 m, 500 m, and 1000 m. Line buffers at the same distances: 100 m, 300 m, 500 m, and 1000 m. Additionally, a catchment scale was used. 2. Geometric Configurations: Point Buffers: These are circular buffers centred around specific points. Line Buffers: These are linear buffers that extend along a line, such as a river or stream. Catchment Areas: These encompass the entire area that drains into a particular water body. Most Influential Buffer Distance on Water Quality The research indicated that the 300 m line buffer was the most sensitive spatial response scale for dissolved heavy metals during the wet season. In contrast, the 1000 m line buffer and catchment scale were identified as the most sensitive in the dry season. Overall, the 100 m point buffer and the catchment scale were noted to have the highest overall explained variance for heavy metals across all media, indicating their significant influence on water quality.
PR5	[38]	Chat Answer: In the research conducted on the Zayandeh-Rud basin, the following buffer distances and geometric configurations were identified: Buffer Distances and Geometric Configurations Buffer Distances (10 km buffer and 15 km buffer) Geometric Configurations: the study utilised both circular and linear configurations for the buffer zones around the monitoring stations. Impact on Water Quality The analysis indicated that the sub-basin scale had the most influential impact on water quality. The results showed a significant positive correlation between land use types and water quality parameters at this scale, suggesting that changes in land use, particularly urban and agricultural expansion, directly affected water quality. In summary, while both buffer distances were used, the sub-basin scale was highlighted as the most effective for understanding the relationship between land use changes and water quality in the Zayandeh-Rud basin.

Note: for clarity, the table comprises solely the responses provided by the CAI, leaving out any supplementary text from the interface.

presents two examples of questions for RO5, including the prompts, references, and responses. The CAI successfully identified various buffer distances (100 m, 300 m, 500 m, and 1000 m) and analysed configurations (points and lines) as noted in ref. [49]. It also provided a consistent explanation, stating that a 300 m line buffer was the most sensitive spatial response scale for dissolved heavy metals during the wet season.

The CAI correctly identified the buffer distances of 10 and 15 km mentioned in [38] and recognised the watershed as the primary spatial unit influencing WQ. However, it erroneously stated that both circular and linear configurations were used, when in fact only circular buffers were implemented. While the study provides graphical representations of these circular buffers that are suitable for readers, the CAI likely faltered due to the absence of explanatory text regarding the use of linear versus circular configurations.

The CAI interpretation of RO5 receives an average score of 3.5. It indicates that while it provides a valuable framework, significant improvements are required for the accurate extraction of key topics. The analysis shows that the prompts effectively identify related text segments, but their effectiveness varies, leading to some misleading responses.

4.5. Key Topic Extraction from Conversational Artificial Intelligence Outputs with Human Validation

4.5.1. Frequently Used Physicochemical Water Quality Parameters

The articles offer valuable insights into water sampling techniques, conservation strategies, and the regulatory frameworks pertaining to the measurement of parameters. Although a complete assessment is not within the remit of this study, it is important to recognise that frequently referencing these practices is beneficial, as it enhances the trust of readers.

In total, 121 WQ parameters were identified, with the ten most employed illustrated in Figure 7: total phosphorus (TP), total nitrogen (TN), pH, dissolved oxygen (DO), electrical conductivity, ammonia nitrogen (AN), chemical oxygen demand (COD), temperature, nitrate nitrogen (NN), and biochemical oxygen demand (BOD). The highest frequency recorded was 34 for TP, while the lowest was 13 for BOD.

TP and TN are the most utilised parameters, appearing 34 and 30 times, respectively. These measurements capture all forms of phosphorus and nitrogen from urban and agricultural sources. Their importance is underscored by esteemed organisations such as [122], which highlight their role in monitoring aquatic systems to ensure compliance with designated uses, shaping policies for pollutant discharge elimination systems, and facilitating the creation of total maximum daily loads for the restoration of non-compliant waters [123].

pH, which measures acidity or alkalinity, appears a total of 29 times. Various natural and anthropogenic factors influence it. They include plant nutrient availability, the effectiveness of pesticides and herbicides, and the breakdown of organic and inorganic matter. Urban activities associated with the chemical, petrochemical, and mining sectors can also impact pH levels [124].

DO is noted 26 times. It indicates the amount of oxygen available to aquatic organisms such as fish and invertebrates. Low levels of dissolved oxygen often indicate contamination and are crucial for assessing WQ, pollution control, and treatment processes [125].

Electrical conductivity is mentioned 22 times. It is known to be significantly correlated with various parameters, including the temperature, pH, alkalinity, total hardness, calcium concentration, total solids, total dissolved solids, chemical oxygen demand, chloride levels, and iron concentration in water. In streams and rivers, conductivity is primarily influenced by the geological characteristics of the surrounding area [126].

AN, which includes un-ionised ammonia and ammonium, appears 20 times. This parameter is crucial for assessing WQ, entering aquatic environments mainly through urban effluent and agricultural runoff. It can generate profound effects on aquatic ecosystems, including the proliferation of cyanobacteria blooms and habitat degradation. Thus, measuring ammonia nitrogen is essential for early warning decision-making activities [127].

COD is mentioned 19 times. This key parameter assesses organic pollution levels in water bodies [128]. Performing rapid and accurate COD detection is vital for daily WQ monitoring, sewage treatment, and pollution source identification [129].

Temperature was noted 17 times. It significantly impacts water chemistry and the health of aquatic organisms, affecting dissolved oxygen levels, the rate of photosynthesis in algae and aquatic plants, and the metabolic rates of various organisms. Temperature also influences the sensitivities of organisms to pollutants, parasites, and diseases, as well as their reproductive timing and migration [130].

NN, cited 15 times, is the end product of nitrification in the nitrogen cycle and can indicate water pollution levels. High nitrate levels in rivers can lead to rapid algal growth and eutrophication [131].

BOD is cited 13 times as an indicator of oxygen consumption by microorganisms, reflecting WQ and pollution levels. Its frequent use highlights its relevance for assessing organic pollution in aquatic systems.

The selection of WQ variables could be influenced by local environmental regulations, data availability, specific research priorities, and resource constraints. It should be noted that recent evidence [132,133] demonstrated that WQ variables involve complex geoenvironmental interactions. These variables could be influenced by a combination of natural processes, such as hydrology and soil composition, as well as anthropogenic activities, including LULC as pollution sources.

The variables are categorised within the extensively researched realm of water pollutants. This is detailed in a comprehensive review that explores the history, models, efforts, and perspectives on WQ indices [134]. The parameters vary in usage frequency in developing WQ indices at the regional and national levels. However, they do not encompass emerging contaminants, which can include pollutants from agriculture and pharmaceuticals [135]. Emerging contaminants in aquatic environments often lack regular monitoring or regulatory oversight, yet they pose significant risks to human health and ecosystems.

4.5.2. Statistical Techniques for Assessing Correlations or Predictive Relationships Between Land Use/Land Cover and Water Quality Variables

The statistical techniques for assessing the correlations and predictive relationships between LULC and WQ variables are described separately for clarity. In practice, researchers often apply these techniques in a complementary manner, combining their strengths to address specific analytical objectives more effectively.

Methodologies frequently highlighted in the literature include Redundancy Analysis (RDA), Principal Component Analysis (PCA), and various correlation techniques, such as linear regression, Spearman’s rank correlation, and Pearson’s correlation coefficient. These observations are consistent with the reviews by [9,11], which underscore the ongoing significance and broad application of these statistical methods. Their simplicity and effectiveness continue to draw in a diverse range of researchers.

RDA is a constrained form of PCA that uses explanatory variables to account for the variation in multiple response variables through multivariate linear regression. It is a versatile method that can be applied with or without the inclusion of landscape metrics, depending on the analytical objectives and the data structure. Ultimately, RDA remains one of the most widely adopted statistical methods for investigating the relationships between LULC and WQ. For instance, studies [49,56,60,93,96] employed RDA to explore the intricate connections between LMTSs and WQ parameters. Refs. [40,53,66] conducted an RDA to assess how LULC proportions affect WQ and to determine the optimal buffer scale.

Furthermore, study [57] applied RDA to assess how TP and TN respond to various explanatory variables within buffers. Refs. [40,95,103] successfully implemented RDA to examine seasonal variations in WQ dynamics.

PCA effectively evaluates the intricate connections between LULC and WQ. The findings of this systematic review support and reinforce that statement. PCA identifies the key parameters that account for the variability in stream WQ [39,45]. A notable instance [73] utilised PCA on WQ indices, employing varimax rotation on standardised component loadings. This method identified the key types of pollution and examined the interactions of WQ indices at each site.

Alternative methods of correlation analysis were applied in recent studies, including the following:

• Chukwuka et al. [90] used linear regression to emphasise the relationship between mining LULC and its impact on WQ. This underscores its application in promoting sustainable mining practices and strategies to mitigate impacts on WQ.
• Stępniewski et al. [75] employed Spearman’s rank correlation analysis to compute correlation coefficients among different types of LULC metrics and various WQ parameters.
• Paná et al. [39] used Pearson’s correlation coefficient to calculate the correlation between physicochemical variables, WQ indices (WQIs), and LULC data at two spatial scales during the summer and winter. While [78] employed the quoted coefficient among twelve WQ indicators and diverse LMTSs.

4.5.3. Buffer Configurations and Their Practical Significance

Studies have mentioned various geometric configurations, including linear, e.g., [40,43,50,62,68], circular, e.g., [38,55,71,78], or both, e.g., [66,85]. Furthermore, spatial domains like catchments, watersheds, and sub-watersheds were incorporated into the analysis of spatial dimensions, e.g., [49,69,82,97,101,104]. The concept of “mixed buffers” is prominent, which involves taking the intersection of the sub-basin and buffers. This approach offered the advantage of producing multiple scales of analysis [88].

Evaluating different distances and geometric configurations clarified how varying distances and shapes affect the influence of distinct WQ variables. For instance, ref. [68] concluded that a buffer width of 200 m was the most effective scale for demonstrating the relationships between landscape patterns and water chemistry, outperforming the results observed at larger scales. Brumberg et al. [98] reported that broad and long riparian forests are ideal for maximising the protection of river WQ and other ecosystem services. However, in landscapes where such conditions are impractical, the 15 m wide riparian forest buffers supported by Costa Rican legislation could still enhance WQ.

The documentation also considered single buffer distances, as evidence suggested it might be optimal for specific queries. For example, ref. [68] demonstrated that a 100 m riparian buffer significantly influenced WQ, particularly in reducing sediment phosphorus concentrations and enhancing nitrogen removal efficiency. In contrast, ref. [105] employed 1000 m to demonstrate how LULC dynamics affect WQ.

The impact of seasonality on buffers was also assessed and there are the following relevant examples:

• Mendivil-García et al. [44] indicated that during the dry season, the coefficients of determination for TP and TN were significantly higher for smaller buffers, suggesting a strong correlation between agricultural coverage and nutrient concentrations. In contrast, the correlation was weaker during the rainy season, with coefficients of determination of less than 0.2.
• Chen et al. [49] revealed that the most sensitive spatial scales for dissolved heavy metals were the 300 m line buffer in the wet season and catchments in the dry season. The most sensitive spatial scales for suspended particulate heavy metals were catchments in the wet season and the 300 m and 500 m line buffers in the dry season.

Buffers have been typically demarcated at several hundred-metre scales; however, various instances analysed relatively large buffers measured in km. In their study, ref. [44] examined buffer distances of 5, 10, 25, and 35 km, finding that the most significant effects on WQ—particularly regarding TN and TP concentrations—came from the buffer areas of 5 km and 10 km. In contrast, ref. [88] determined that a 25 km buffer had the most pronounced impact on TN and TP indicators, while [90] focused on the 2 km and 5 km buffers to evaluate the implications of mining activities.

The analysis of the literature suggests that there is no consensus about BZSs’ meaning because the concept sometimes refers to generic spatial units of study or areas designated for specific management purposes.

4.6. Findings in Context: Limitations and Implications for Policy and Practice

The evidence presented, despite its limitations and biases, can support strategies to improve WQ in buffers and their watersheds.

A comparison of this research with similar studies demonstrates that previous work covered a wider range of topics regarding the impact of LULC on WQ. For instance, Cheng et al. [9] analysed the effects of different LULC types, including agricultural, urban, and forest areas. They also evaluated the potential of prominent RS missions from 1998 to 2021. Wu et al. [10] addressed challenges related to water security and quality, specifically focusing on the role of riparian BZSs in managing WQ and mitigating agricultural non-point source pollution. Locke [11] explored the distinctions between temporal and spatial perspectives, defining and classifying natural vegetation while underscoring the importance of landscape configuration, location, scale, and potentially confounding environmental factors. Shi et al. [12] synthesised evidence on landscape–WQ studies, highlighting the impacts of landscape changes on WQ, identifying moderators of this relationship, and assessing WQ responses to different LULC types.

Conversely, this systematic approach targeted specific evidence to meet RO1, RO2, RO3, RO4, and RO5. It complements existing knowledge by identifying current trends related to LULC maps, key WQ parameters, and standard methodologies such as STATs and buffer configurations. Cited contributions clearly make this review different from previous experiences but complement them.

The implications of this review extend beyond academic relevance, offering practical insights for policy, land management, and environmental planning. For instance, RO1 outcomes highlight the need for accessible and scientifically validated LULC datasets, which can facilitate consistent decision-making across buffers and watersheds. The prevalence of specific LMTS (RO2) reports helps identify the most robust indicators for monitoring LULC patterns, fragmentation, and connectivity, directly supporting the planning of buffers that may function as ecological corridors or protected areas.

Additionally, the consistent use of specific physicochemical WQ parameters (RO3) reveals the aspects most sensitive to land use pressures, suggesting these should be prioritised in monitoring and management strategies. The emphasis on statistical approaches (RO4) underscores the value of practical techniques that have stood the test of time. Finally, RO5 findings suggest the optimal buffer distance and geometric configuration remains a complex and context-dependent task, influenced by the characteristics of surrounding LULC, the specific WQ parameters being assessed, and seasonal dynamics that affect ecological and hydrological processes.

Based on the findings from this research, concrete and actionable initiatives could be implemented in various ways, as illustrated by the following examples:

• LULC instrumental planning. The identified relationships between LULC and WQ can guide land use zoning and territorial planning. For instance, local authorities could restrict urban expansion in sensitive headwater zones or designate agricultural areas where best management practices are mandatory. At a regional scale, planning authorities could integrate WQ indicators into watershed development plans to ensure that economic growth aligns with ecological sustainability.
• Restoration and conservation programmes. Evidence on the benefits of riparian vegetation and forest cover suggests prioritising the reforestation of degraded streambanks, the establishment of constructed wetlands for nutrient retention, and the recovery of floodplains to reduce sediment loads. These measures can be strategically implemented in hotspots where LULC changes have the most substantial adverse effects on WQ.
• BZS delineation. Rather than applying a uniform buffer width, the results support a more tailored approach, where buffer distances are defined according to LULC intensity and the specific WQ parameter of concern (e.g., wider buffers in agricultural areas to reduce nitrogen and phosphorus runoff). This adaptive delineation could be formally incorporated into watershed management plans and environmental regulations.

The application of the CAI demonstrated a mixed performance. On one hand, it proved accurate in identifying variables related to WQ, facilitating a more rapid and structured synthesis of findings. On the other hand, its interpretation of statistical aspects and spatial buffer configurations was less precise, necessitating additional human intervention to ensure completeness.

5. Conclusions and Recommendations

This systematic review examined contemporary trends in the application of LULC mapping for the evaluation of WQ in streams and buffer environments. This review demonstrated the significant advantages of integrating NLP with human analytical techniques.

Given the evident regional disparities in research coverage, particularly in underrepresented areas of the Global South, future reviews—especially those with a local focus—should incorporate the literature from languages other than English and explore reliable sources beyond merely peer-reviewed journals.

Future research, whether in the form of systematic, exhaustive, or comprehensive reviews, should examine the impact of LULC from the perspective of lentic, marine, or groundwater bodies. The dynamics of the LULC-WQ nexus in these systems differ significantly from those analysed in streams or rivers. Moreover, as WQ simulation programmes and hydrographic modelling represent evolving areas of inquiry driven by advancements in technology—which are further boosted by RS and geographic information systems—synthesising the relevant evidence becomes a priority.

Current research often neglects advances such as multisource RS, recent classifiers, deep learning algorithms, and cloud-based platforms. Future investigations should integrate these tools and exploit the potential of geo big data, while also recognising that disparities in access to infrastructure, funding, and expertise may affect their adoption. Facilitating the transfer of advanced RS methods, such as deep learning classification and cloud computing, to scientific communities with less specialisation in RS could be a transformative journey. This transition, while complex, opens doors to new possibilities and innovations, requiring specific technical expertise and computing resources that may not be universally accessible. To empower these communities, reputable academic and spatial agencies must be dedicated to disseminating massive open online courses and tutorials, providing invaluable opportunities to acquire and apply essential knowledge practically.

To improve the reliability of these findings, subsequent research exploring the effects of LULC on WQ should systematically incorporate critical accuracy metrics, including OA and the Kappa index. Additionally, the use of class-level metrics, such as user and producer accuracy, would significantly enhance the depth and validity of the analysis.

Effective ecological management requires the consideration of spatial differences in LULC across river reaches, as well as varying land management strategies within the same segment. This approach will aid in establishing adequate buffers.

QDAS effectively manages and analyses extensive bibliographic references. CAI has demonstrated value in expediting systematic reviews, particularly in identifying WQ parameters, but it remains limited in handling the requested statistical content. Continued user training will be key to optimising its application.

Researchers are encouraged to explore alternative or complementary CAI platforms—such as ChatPDF, Humata.ai, or domain-specific tools—and, where possible, to develop tailored solutions that address specific research needs. In summary, while CAI and QDAS provide promising support for systematic reviews, their practical use requires careful integration with human expertise.

As CAI is a relatively new field, it is anticipated that technologies will continue to improve and standardised best practices will emerge. Therefore, the scientific community should remain mindful of these developments while conducting critical human-based and ethical analyses.