There is currently no operational method of managing irrigation in Australia's sugar industry on the basis of systematic, direct monitoring of sugar plant physiology. Satellite remote sensing systems, having come a long way in the past 10 years now offer the potential to apply the current ground-based 'FAO' or 'crop coefficient (K_c)' approach in a way that offers a synoptic view of crop water status across fields. In particular, multi-constellation satellite remote sensing, utilising a combination of freely available Landsat and Sentinel 2 imagery, supplemented by paid-for imagery from other existing satellite systems is capable of providing the necessary spatial resolution and spectral bands and revisit frequency. The significant correlations observed between K_c and spectral vegetation indices (VIs), such as the widely used normalised difference vegetation index (NDVI) in numerous other crops bodes well for the detection and quantification of the spatial difference in evapotranspiration (ET_c) in sugar which is necessary for irrigation scheduling algorithms. Whilst the NDVI may not serve as the appropriate index for sugarcane, given the potential of the NDVI to saturate at the high leaf area index observed in fully developed cane canopies, other VIs such as the Green-NDVI (GNDVI) may provide the response required. In practise, with knowledge of an appropriate K_c-VI relationship, K_c obtained from time-series (weekly) remotely sensed data, integrated with local agrometeorological data to provide ET_o, would provide estimates of ET_c from which site-specific irrigated water requirements (IWR) could be estimated. The use of UAVs equipped with multispectral sensors, even active optical sensors (AOS), to 'fill the gaps' in optical data acquisition due to cloud cover is conceivable. Cross calibration of any passive imaging system, as with the multi-constellation satellite data is essential. The use of radar images (microwave remote sensing) (for example, Sentinel 1&2 C-SAR, 5m) offers all weather, day-and-night capabilities although further work is necessary to understand the link between the radar back scatter, which is responding to surface texture, and evapotranspiration (and K_c). Further R&D in ascertaining the K_c-VI relationships during crop growth is necessary, as is the testing of multi-sensor cross-calibration and the relationship between radar remote sensing and K_c. Existing irrigation advisory delivery systems in Australia such as IrriSAT should be investigated for their applicability to the sugar industry. The estimated season cost to a user for a sugarcane irrigation advisory service in Australia, based on the use of data from existing optical satellite imaging systems and utilising the K_c approach, is likely to be of the order of US$2-3/ha.

The performance of a multi-view machine vision method was documented at an orchard level, relative to packhouse count. High repeatability was achieved in night-time imaging, with an absolute percentage error of 2% or less. Canopy architecture impacted performance, with reasonable estimates achieved on hedge, single leader and conventional systems (3.4, 5.0, and 8.2 average percentage error, respectively) while fruit load of trellised orchards was over-estimated (at 25.2 average percentage error). Yield estimations were made for multiple orchards via: (i) human count of fruit load on ~5% of trees (FARM), (ii) human count of 18 trees randomly selected within three NDVI stratifications (CAL), (iii) multi-view counts (MV-Raw) and (iv) multi-view corrected for occluded fruit using manual counts of CAL trees (MV-CAL). Across the nine orchards for which results for all methods were available, the FARM, CAL, MV-Raw and MV-CAL methods achieved an average percentage error on packhouse counts of 26, 13, 11 and 17%, with SD of 11, 8, 11 and 9%, respectively, in the 2019–2020 season. The absolute percentage error of the MV-Raw estimates was 10% or less in 15 of the 20 orchards assessed. Greater error in load estimation occurred in the 2020–2021 season due to the time-spread of flowering. Use cases for the tree level data on fruit load was explored in context of fruit load density maps to inform early harvesting and to interpret crop damage, and tree frequency distributions based on fruit load per tree.

Globally, the production and trade of fruits and nuts from horticultural tree crops (HTCs) is increasing due to greater demand from a growing world population. Among those HTCs, is the mango, venerated as the “king of fruits” due to its nutritional, health and the economic benefits it provides to both developed and developing nations. Its production and trade have consistently risen since the early 1960s, when official reporting begun. Global production increased from 10.9 million tons in 1961 to over 57 million tons in 2021, representing a 422% increase. According to the Food and Agricultural Organization of the United Nation (FAO), mango contributed USD 0.6 million to approximately USD 3.7 billion from 1961 to 2021 in export value to the global economy.

With increasing food demand, there is the need for increased production, optimizing efficiencies and minimizing environmental impact through the more judicious use of crop inputs. Part of this solution is the development of technologies and analytics that can more accurately and efficiently measure the spatial and temporal variability in tree health, timings of key phenological stages that dictate key management practices and production (yield and quality). The outcomes of these applications also assist with improved decision making around harvest planning and logistics, minimizing potential food wastage along the value chain, market access and forward selling. The current commercial practice for measuring these key parameters in tree crops, including mango, is predominantly by in-field within season assessment which is costly, time and labour intensive, can be inaccurate due to a nonrepresentative area of the orchard being evaluated and human subjectivity.

A review of prior literature identified some recent technological advancements such as the use of weather parameters ‘Growing Degree Days’ for determining tree growth phase and fruit maturation; proximal sensing/ machine vision and the targeted manual fruit counts of individual trees for calibrating remotely sensed imagery. However, these methods are costly, time and labour intensive. Additionally, the approaches can lack spatial granularity, scalability, commercial readiness and provide accuracies fairly similar to current commercial practices. From the publications reviewed across many horticulture and agricultural crops, remote sensing (RS) and associated cutting edge analytics (e.g. Machine learning (ML)) presents as the most likely technology to improve current management and forecasting practices in mango. However, a large knowledge gap still remains.

This study sought to address this knowledge gap by undertaking, four key objectives:

1. Assess the potential of Sentinel-2 satellite data derived vegetation indices (VIs) in distinguishing phenological stages of mango (Chapter 2)

2. Assess the potential of Sentinel-1 satellite data in distinguishing mango phenology and investigating its relationship with weather variables (Chapter 3)

3. Explore the relationship between very high-resolution satellite imagery data and fruit count for predicting mango yield at multiple scales (Chapter 4), and finally

4. Integrate time series remote sensing and weather variables for mango yield prediction using a machine learning approach (Chapter 5).

The study was conducted in commercial mango orchards in both Ghana and Australia covering the period 2015 to 2022 using RS data obtained from platforms such as Sentinel-1 (S1), Sentinel-2 (S2), Landsat-8 and WorldView-2 (WV2) and WorldView-3 (WV3) to derive VIs. A number of Statistical and ML approaches including Linear regression (LR), Random Forest (RF), Support Vector Regression (SVR), eXtreme Gradient Boosting (XGBOOST), Ridge, Least Absolute Shrinkage and Selection Operators (LASSO) and Partial Least Square (PLSR) regressions were employed at various stages throughout the study

Four publications were produced during the study, with the key findings of each publication presented as follows:

• The S2-derived Enhanced Vegetation Index (EVI) was identified as the index that best distinguished five phenological stages of mango (Flowering/Fruitset (F/FS), Fruit Development (FRD), Maturity/Harvesting (M/H), Flushing (FLU) and Dormancy (D)) of four mango farms in Ghana.

• S1-derived radar VI was identified to be responsive for distinguishing three phenological stages (Start of Season (SoS), Peak of Season (PoS) and End of Season (EoS)) from a mango farm in Ghana. These stages align well with three of the key phenological stages (F/FS, M/H and D) retrieved in the optical data (S2) experiment above. It was also established that although weather is known to influence growth and yield of HTCs, for the weather conditions of the study area, its influence on phenology was marginal.

• The evaluation of 24 WV3-derived VIs from individual tree canopies and associated fruit counts collected across many locations, seasons and cultivars (n = 1958), identified no consistent generic relationship between the predictor (24 VIs) and response (fruit count) variables at the individual orchard level. The subsequent modelling of all composite data through ML algorithms, identified the RF-based yield prediction accuracy was better at the farm level than the individual tree level with percentage root mean square error (PRMSE) of 10.1% and 26.5% respectively, for the combined model (i.e. a model trained on all cultivars, locations and seasons data). The potential of developing an ML-based yield variability map at the individual tree level to support precision agriculture was demonstrated.

• An RF-based time series model is capable of predicting block and farm level mango yield around 3 - 5 months ahead of the commencement of the commercial harvest season. The block level combined RS/weather-based RF model for 2021 produced the best result (mean absolute error (MAE) = 2.9 t/ha), marginally better than the RS only RF model (MAE = 3.4 t/ha). The farm level model error (FLEM) was generally lower than the block level model error, for both the combined RS/weather-based RF model (farm = 3.7%, block (NMAE) = 33.6% for 2021) and the RS-based model (farm = 4.3%, block = 38.4% for 2021). The errors thus, ranged from 3.7% to 82.7% and 28.7% to 70.7% at the farm and block levels respectively, for the RS/weather-based RF model across the 8-year time series. Factors such as irregular bearing and data associated limitations were possible causes of errors in the study. The study demonstrated the ability to improve yield prediction accuracy from a finer (e.g. block level) to coarser (e.g. farm level) scales as positive (overprediction) and negative (underprediction) errors tend to cancel out.

Overall, this study demonstrated the potential of integrating RS and weather variables for accurate mango yield prediction. Nevertheless, challenges including a lack of extensive farm level standardized data and the high cost of high-resolution imagery exist. Additionally, whilst these methodologies did demonstrate some benefit over existing practice, further validation of these methodologies is required over more growing locations, cultivars and seasons. This also includes the extrapolation of models at multiple scales, particularly regional and national levels (which were beyond the scope of this study). Furthermore, future research could explore the potential of this method to produce robust estimates in other perennial tree crops.