Introductory Chapter: Present and Future of Artificial Intelligence in Grasslands Conservation

1. Introduction

The grasslands entail any ground cover encompassing grasses as a predominant vegetation, whereas shrubs and tree cover generally tend to prevail on less than 10% of that area [1, 2]. Recently, grasslands conservation and development have attained the attention of researchers and policy makers owing to their strategic pertinence in terms of carbon sink, watershed areas, feed source for livestock etc. Fundamentally, the concept of grasslands conservation involves those initiatives which ensure protection of grassland ecosystems and sustainable provision of ecosystem services through scientific maintenance of their ecological integrity and preventing the loss of biodiversity. Besides this, there is another concept of grasslands development which encompass different anthropogenic interferences (e.g., construction of roads, dams, wild-life sanctuaries, amusement parks etc.) intended for diversifying and multiplying the ecosystems services offered by grassland areas primarily through enhancement of their agricultural usage [3]. Traditional grassland ecosystem monitoring has mainly relied on field surveys, however these are being increasingly enriched and developed towards a large range, high spatio-temporal resolution and high-precision directions. The conservationists and researchers call for innovative solutions to effectively manage a variety of anthropogenic interferences and handle environmental problems that have recently seriously threatened the ecosystem functioning of grasslands.

Grasslands conservation requisites the exploration of their productive potential in order to plan grazing and reseeding activities [1, 2]. Currently, approaches employed for grasslands monitoring (e.g., visual assessments) are labour demanding, costly and, therefore have remained inadequate in terms of practical use. Additionally, nutritional value of native grass species has remained neglected and un-monitored since their quality assessment requires intricate laboratory analyses, involving huge expenditures and technical expertise [4]. Alternatively, artificial intelligence (AI) based tools might be developed and optimized as viable, reliable and cost-effective methods to monitor and assess grassland production potential under changing pedo-climatic scenarios [5]. Moreover, digitalized monitoring using automated tools might assist in recognizing native grass species, invasive species and analysing soil fertility status for formulating effective management and conservation options. For instance, unmanned aircraft systems (UAS) integrating numerous types of camera and sensors hold potential to precisely collect scattered spatial and temporal information having remarkably high-resolution in the visible and infrared spectrum [6]. Likewise, AI, deep learning (DL) and machine learning (ML) could aid to construct 2D and 3D models [7] for establishing grass species composition, total vegetation cover, barren patches, and nutritional quality of grass species. Furthermore, ML and AI algorithms may be employed for localizing the grass species from the images and sensors supplied data. Using these approaches, grass species can be localized down to centimetre scale and allows maps construction to illustrate grass cover distribution on large swathes of grasslands.

This chapter briefly introduces fundamentals and application types of artificial intelligence with special emphasis on deep learning, drone’s types and their utilization potential for grasslands monitoring and conservation. Moreover, different challenges that could emerge while using artificial intelligence tools for grasslands conservation and development have also been objectively highlighted.

2. Artificial intelligence (AI)

The AI entails an amalgamation of technologies having human-like cognitive capacities which enable them to precisely learn, accurately perform, and effectively make decisions. It is fundamentally a software which might initiate logical reasoning, learn from patterns, and resultantly can solve intricate problems in a highly rationale manner [8, 9]. Interestingly, AI is typically defined in a specific computer science method’s context that it uses such as machine learning (ML), reinforcement learning (RL), and deep learning (DL). The AI software can also be physically implemented in the form of humanoid robots, autonomous cars and robotic hands. The presence of AI is felt wherever any type of machine is performing a wide range of tasks by utilizing reasonably high level of independent intelligence that is typically practiced by humans [8, 10]. In comparison to natural intelligence of humans and other biological organisms, the machine intelligence is different in the sense that it is artificially created and digital in nature and functioning. In recent generation AI has emerged as one of the most exciting research area owing to its unprecedented potential to solve the intricate problems particularly to reduce labour use and ensure efficient resource utilization [8, 9, 11]. In agriculture, a wide range of AI approaches have been suggested including robotics, DL, ML, neural network (NN), support vector machine (SVM), random forest etc. Figure 1 illustrates AI applications especially cognitive related applications (data fusion and mining, artificial neural networks etc.), robotics applications (unmanned aerial vehicles or drones, swarm robots etc.) and natural interface concerning applications (virtual and augmented reality, computer vision etc.).

Figure 1.

Different types of artificial intelligence applications especially cognitive applications (data fusion and mining, artificial neural networks etc.), robotics applications (unmanned aerial vehicles or drones, swarm robots etc.) and natural interface applications (virtual and augmented reality, computer vision etc.).

Broadly, intelligent systems are classified into four categories including (і) systems having humans like cognitive potential, (іі) systems capable to act like humans, (ііі) systems that rationally think to solve any problem, and (іν) systems which act rationally to solve intricate challenges [9, 12]. This classification of intelligent systems is based on their potential to think and act while their success is measured in comparison to human rationality and performance. Thus, AI systems hold potential to assist in grasslands conservation as these can be effectively used for manipulation and long-term storage of data regarding grass cover and composition along with soil condition and pest record. Data manipulation entails potential to deduce and infer novel knowledge from existing information and data set. Additionally, these AI systems can precisely record data through vigorous acquisition of evaluation parameters at inaccessible altitudes of grasslands along with data analyses and representation through interpretation of the acquired knowledge. Among AI techniques, artificial neural networks (ANN) has emerged as one the most strategic technique that is developed by interconnecting nodes having potential to perform intricate functions as human brain [10, 13, 14]. This technique has been successfully employed in agriculture for estimating and forecasting methane emission from fields and soil moisture contents. Besides, scaled conjugate gradient and Quasi-Newton based neural network algorithms have been employed to estimate soil traits and environmental variables impact and predict dynamics of soil moisture content on hourly basis [12, 14]. However, focus of this study is robotics, deep learning and drones for their potential utilization in grasslands conservation and development.

3. Robotics

Robotics a branch of technology aims to design and construct robots for different applications and operations. Robots are basically machine resembling humans having potential to replicate actions typically performed by humans [8, 9]. There have been industrial robots and mobile robots during last many decades and now service robots have been focussed in order to develop their closer interaction with humans to fulfil their emerging needs. Few pertinent examples of service robots include medical robots, underwater robots, field robots, rehabilitation robots, construction robots and humanoid robots [9]. Among AI approaches, robots are currently most widely used in agriculture after drones to assist in many operations including visual detection, weeding, spray application, and harvesting having flexibility and adjustment options to match the requirement of various tasks. Such robots are typically two-wheeled entailing a mobile base to execute a spewing mechanism to control the movements of the robot through a wireless tool. Additionally, robots might be fitted with high resolution cameras to monitor the growth conditions of native grass species and monitor the effects of abiotic stresses (drought, heat, chilling, salinity, heavy metal toxicity, soil compaction, water logging etc.). Moreover, this monitoring can effectively reveal and detect the presence and extent of pest attack in different areas of grasslands. Similarly, robots can be equipped to collect soil samples for determining the physico-chemical characteristics of the grasslands soils at varying altitudes which are typically inaccessible for humans. Moreover, robots can perform the function of surveillance for reporting over-grazing and invasion of exotic species in the grasslands which in the longer run results in the deterioration of ecosystem functioning of the grasslands. However, research and practical implementation of robots based surveillance programs are still awaited for grasslands monitoring and conservation.

4. Deep learning (DL)

The DL technology is relevant to AI, ML and data science (DS) with advanced analytics and recently it has emerged as a focussed research area in computer science and intelligent computing. Generally, AI is the name of incorporating human intelligence and behaviour to different machines, set of machines or complex machine systems [13], while ML represents learning method from data set or experience, which automates the building of analytical models. Likewise, DL also entails learning from data set whereby computation is processed through multi-layer neural networks [7, 13, 15]. In DL, the term deep encompasses the concept of multiple stages or levels through which processing of data are performed in order to build a data-driven model. Therefore, DL is a technique that trains multiple computers enabling them to process information in such a way which mimics human neural processes. It is fundamentally a ML technique that instructs computers to learn by example from large data sets and architectures of neural networks. The DL technique has been successfully employed to teach computers that imitates the way humans gain and process intricate information and knowledge [13].

The original idea of DL was envisaged in 1943 [12] and it was meant for creating a model entailing a single biological neuron and its progressive development to link numerous individual neurons together to build an ANN. Recently, deep neural networks (DNN) have superseded the ANN by virtue of comparatively larger number of layers and multiple data sets. It has been enabled to separate signal of interest from noise in data set by utilizing algorithms including backpropagation in order to optimize parameters of interest in a single layer from the previous layer [13].

Recently, DL based methods have been developed to process the data set of ultra-high-resolution images which might be utilized for monitoring the grasslands features like vegetation cover and extent of exotic species invasion. Particularly, deep CNNs (convolutional neural networks) have recently revolutionized the image processing and their interpretation through remarkable improvement in specified object detection and precise classification of assigned tasks [13, 15]. Interestingly, deep CNN is basically a network which is comprised of many layers to take image’s pixels as input data to predict interpretation as output. The underlying mechanism is CNNs primarily apply different learned filters (thousands in number) to the image’s regions internally and finally combines those to find out targeted information [15]. Importantly, DL uses end-to-end approach which allows targeted features automatic detection without human interaction owing to its capability to learn and detect the desired features. Recently, such methods have been successfully introduced for the detection and counting of plants and animals in varying ecological context and this feature might be utilized to recognize native grass species cover in grasslands in order to formulate the organize conservation initiatives.

5. Drones

Drones are basically airborne self-propelled devices having a programmable regulator for controlling and regulating their movement without any on-board pilot. Since last decade, drones also known as unmanned aerial systems (UAS), remotely piloted aircraft systems (RPAS) or unmanned aerial vehicles (UAV) have attracted the attention of researchers in civilian and scientific spheres, because of their potential to usher a new era of precision agriculture, remote sensing and environment studies [16, 17]. The UAV can provide images of high resolution of natural phenomena at a relatively low-cost, risk-free, rapid and systematic manner [18, 19]. For these reasons, UAV hold potential to become a major trend in grasslands and wildlife related research. Drones are classified in various ways-by size, range, endurance and carrying capacity. Figure 2 portraits the classification of different types of drones based on their size, range, take-off weight, wing type, power source, assembling requirement and nature of use or applications (monitoring or logistic). However, wing based categorization is the simplest classification especially fixed wing drones are capable to carry heavy loads over comparatively longer distances while working just like airplanes [16, 19, 20, 21, 22].

Figure 2.

Classification of different types of drones based on their size, range, take-off weight, wing type, power source, assembling requirement and nature of use or applications (monitoring or logistic).

Likewise, ready to fly drones have higher take-off weight potential and thus, these hold bright perspectives to perform over-seeding operation on vast swathes of grasslands. Contrarily, small sized (fixed-wing and rotary-wing) drones might find their use in generating high resolution videos and still photography of grasslands situated on varying altitudes. Such types of UAVs equipped with multispectral sensors and lightweight cameras can potentially deliver real-time professional mapping at a fraction of the cost involved in previously human-employed photogrammetric techniques. Additionally, compact thermal vision cameras might also be installed on medium size UAV along with hyperspectral sensors and laser scanners [19, 21] for vegetation studies in grasslands and forests. Moreover, drones can be equipped with sensing devices [18, 21] for recording distinct environmental attributes (temperature, relative humidity, precipitation or air pollution) of far-flung grasslands. Moreover, large size aerial platforms having capability to lift heavier payloads might represent one of the most economical solution for sampling of grasslands soils or deliver grass seeds in degraded patches.

The UAV’s success in performing grassland conservation operations can be partially predicted by keeping in view their great flexibility to carry different types of cameras, sensors, lasers and other devices. The scope of operations (visual observance, species recognition, grass composition assessment, over-seeding, nutrients application, invasive species intrusion assessment etc.) might determine the optimum combination of airborne devices and payload. These UAVs can assist in accurate retrieval of grassland traits to formulate conservation plans by adjusting grazing frequency and intensity. Optical imaging systems might be carried by airborne devices as payload which might serve as an alternate of non-destructive methods for nutritional quality assessment of grass species. Similarly, vegetation spectral response by employing visible and near-infrared wavelengths can generate useful information on vegetation composition in grasslands. Moreover, by combining multivariate modelling with spectral measurements can reveal complex relationship between grass traits and canopy reflectance. The grasses canopy’s spectral response in the visible and near-infrared (NIR) regions (between 450 and 900 nm wavelength) can unveil biophysical and biochemical traits such as canopy cover, chlorophyll content, primary nutrients (nitrogen, phosphorous and potassium) contents of grass species. Furthermore, UAV borne imaging systems entail capability to provide spatially continuous information, which is much more advantageous in comparison to point-based measurements for exploring spatial patterns of grasslands monitoring parameters. Drones can be equipped to estimate grasses traits based on canopy spectral measurements in the optical domain that is generally classified as empirical and physically-based approaches. Empirical approaches rely on large data set to fit prediction models, whereas physical approaches need adequate parametrization of a Radiative Transfer Model (RTM) in order to translate spectral information into grass traits. The RTM-based frameworks have greater potential for generalization however performance could be suboptimal particularly in grasslands having multiple grass species. Contrastingly, empirical models have better adaptability for complex datasets however, these have limited applications. Thus, quick and risk-free evaluation favours deployment of UAV borne optical imagery, while empirical models might be utilized as an initial option to characterize grasslands.

Traditional assessment methods (survey expeditions by technical staff or researchers) generally focus on recording data of biomass cover seasonally or occasionally which limit drawing accurate conclusions for grasslands conservation and development owing to small data sets. However, grass species in grasslands need frequent monitoring owing to changing climate scenarios, wild fires, uncontrolled intensity of grazing by non-concerned ranchers, deteriorated soil conditions, invasive plant species etc. There is dire need to develop a perennial monitoring systems for grasslands that has potential to cope with rapidly varying climatic conditions because such frequent data acquisition is vital for making informed decisions. Therefore, prediction models relying on limited spectral datasets are always prone to errors while such uncertainties can be effectively dealt with the use of aerial borne platforms equipped with high resolution cameras and state-of-the-art sensor systems.

6. Challenges and future perspectives

The AI based models hold advantages of being flexible and adaptive, however their application for monitoring grasslands can be intricate and opaque. Nonetheless, these opaque models may become computationally expensive (exponentially) with higher levels of data complexity. Despite all advantages, the ANNs are extremely data intensive and requisite time-consuming parameter tuning approaches to process information [9, 10]. These limitations of ANNs have led to the development of easier-to-train algorithms including decision trees (DTs), support vector machines (SVMs), and random forests (RFs), however their application for grasslands monitoring still awaits target-specific research. In addition, there is need to create trust on AI based forecasts along with developing ethical and moral standards of their application in grasslands conservation and development. To achieve these goals, detailed explanations of AI decisions will be necessary in order to provide insights pertaining to the rationale the AI model use to draw conclusions and make forecasts. Additionally, frequent trainings and expertise would be needed to increase the interpretability (the propensity of humans to interpret and understand AI algorithms results). The trust of stakeholders on AI based forecast can be increased through robust explain-ability entailing strong, easy to comprehend and user-friendly justifications on all the decisions and predictions produced by the AI systems. For utilization in grasslands conservation and development, future research need to assess explainable AI types such as (і) opaque systems revealing no reasoning about their algorithmic architecture, (іі) interpretable systems offering mathematical analyses of their algorithmic mechanisms, (ііі) systems relying on symbols such as text and visualizations to allow users to understand how algorithm perform conclusions, and (іν) explainable systems that perform automated reasoning to craft concise explanations of their algorithmic mechanisms with minimum human interferences [14].

The assigning of clear responsibility and subsequent accountability of AI based forecasts and decisions continues to remain one of the biggest challenge and same has been realized for their utilization in grasslands conservation. There must be concise responsibility assignment mechanism before operationalization of AI techniques in grasslands in order to avoid potential accidents, and misleading predictions. In addition, there is need to overcome moral proxy problem associated with AI system because these have been designed by humans for serving human ends and to make decisions on their behalf, but these are not advanced enough so far to differentiate morally right or wrong decisions and processes.

Moreover, future research needs to address challenges like AI based model’s biasing of results owing to non-representative data, narrow applicability due to non-inclusive usage contexts and their potential failure to differentiate among grass species. In grasslands conservation and development, preventing bias (tendency to learn from a preferred data pattern instead of actual data distribution) and pursuing inclusiveness and fairness in AI based approaches are bound to determine the human-machine partnerships in future. Interestingly, an AI system is generally useful as per assumptions of algorithm, however it can only be expected to precisely forecast decisions based on recorded observations, whereas it make inaccurate predictions in unprecedented, novel and unexpected situations. Likewise, rethinking regarding who will own, store, process and control grasslands data is also a challenge which needs redressal to prevent disputes of ownership, privacy as well as security risks.

7. Conclusions

Owing to climate change, human interferences and ecological pressures, it is high time to implement novel solutions for conserving the natural and semi-natural grasslands. These goals could be effectively achieved through deployment of AI based solution. For instance, use of drones and robots merit to perform conservation actions and reinforce effective management, but multidisciplinary research must resolve the operational and analytical shortcomings that undermine the prospects for their integration in grasslands conservation and development strategies. Future usefulness of AI based solutions will depend on accuracy and reliability of recommendations forecasts made by AI systems. This requires research based development of designs and usability of AI systems involving greater transparency and explainability. Additionally, stakeholders need to focus on governance and outcome based AI approaches involving formal sectors (government or corporate oversight) as well as informal sector (ranchers, tourists etc.). Cantering ranchers and other stakeholders during AI model designing and affixing clear responsibilities along with establishment of a chain of accountability regarding AI decisions may serve as a starting point of AI based solutions for conserving grasslands. Moreover, digitalization of grasslands monitoring operation might increase algorithmic bias risks which necessitate algorithmic objectivity on which AI models thrive and interpret the recorded information. Finally, initially soft compliance must be progressively proceeding to stringent laws for preventing abuse of data and AI systems. However, creative and critical thinking are required to create incentives and establish institutions for ranchers and other human population residing in close vicinity of grasslands to engage them in AI based grassland conservation solutions.