Explainable machine vision techniques for cattle disease detection with deep transfer learning

Abstract

Cattle farming is a cornerstone of global agriculture, playing a pivotal role in food security, economic stability, and the livelihoods of millions of people worldwide. In many regions, cattle provide essential products such as milk, meat, leather, and other by-products critical to human sustenance and commerce [26]. However, the industry faces a major threat from diseases such as Foot and Mouth Disease (FMD). These diseases significantly reduce productivity, compromise cattle health, and pose risks to public health through zoonotic transmissions [17], [22]. The impact of these diseases is far-reaching. Cattle diseases result in decreased milk and meat production, poor reproductive performance, and higher mortality rates, leading to substantial financial losses for farmers and the agricultural economy [31]. Cattle in poor health have a diminished ability to work or produce, adversely affecting farm operations [24]. In addition, some diseases pose zoonotic risks, which can directly threaten human health if left untreated or mismanaged [12]. According to [30], traditional methods for cattle disease detection largely rely on manual inspection by veterinary experts. While effective in controlled settings, these methods have several limitations, particularly in cases where many farmers, especially in rural and under-served areas, cannot afford or access veterinary expertise promptly [10]. Manual detection often involves physical examination, laboratory testing, or other diagnostic procedures that are time-consuming [27]. Variability in human judgment and environmental conditions can lead to misdiagnoses or delayed interventions [29]. Moreover, the expenses associated with routine inspections and specialized diagnostic tools are often prohibitive for small-scale farmers [30]. [33]In the last decade, technological innovations have revolutionized cattle disease detection where artificial intelligence (AI), machine learning (ML), and deep learning (DL) techniques have introduced automated, scalable, and highly accurate diagnostic solutions. For instance, the use of Convolutional Neural Networks (CNNs) in diagnosing diseases and bovine respiratory disease has shown remarkable success in processing image-based data [33]. Additionally, wearable devices, cameras and Internet of Things (IoT) sensors have enabled continuous monitoring of physiological parameters, allowing early detection of disease in cattle and preventive interventions [23]. 2However, existing AI-based models for cattle disease prediction face notable gaps that hinder their practical adoption and widespread use. One of the most significant challenges is the lack of transparency in their decision-making processes. Many deep learning models, including Convolutional Neural Networks (CNNs), operate as ”black boxes,” meaning that their internal computations and predictions are not easily understandable by end-users, including veterinarians and farmers. This opacity creates skepticism and limits the trust needed for these technologies to be integrated into day-to-day cattle management practices. Veterinarians and farmers often require a clear understanding of why a specific diagnosis or prediction is made to ensure its validity and take appropriate action. Without this interpretability, users may hesitate to rely on such systems, particularly when the predictions deviate from their expectations or experience. While CNNs and other deep learning approaches have shown impressive performance in analyzing visual data, such as images of diseased cattle, they tend to focus exclusively on this single modality. This reliance on visual data alone ignores other critical factors, such as environmental conditions (e.g. temperature, humidity, pasture quality) and historical health records (e.g., vaccination history, past illnesses). These complementary data sources often play an important role in disease progression and detection accuracy. For example, outbreaks of Foot and Mouth Disease can be correlated with specific climatic conditions, or a history of mastitis can in crease susceptibility to other infections. By not incorporating these contextual and historical factors, existing models can miss subtle but essential patterns, leading to sub optimal predictions and interventions. Additionally, most current models are designed to detect the presence of diseases but do not go beyond simple classification or detection. They lack the capability to provide actionable insights, such as highlighting key features in an image that influenced the prediction. For instance, a model may predict that an image shows a cow suffering from foot and mouth disease, but it might not specify the exact visual symptoms (e.g., swelling) that led to this conclusion. This limitation reduces the interpretability of the models and makes it challenging for veterinarians to validate the predictions or communicate the results effectively to farmers. Moreover, without actionable insights, farmers and veterinarians are left without clear guidance on how to prioritize their interventions or prevent further disease spread, limiting the practical utility of these systems.