Benchmarking new methods for estimation of quantity and harvest timing of the mango crop: Dataset

A forward estimate of mango fruit harvest volume and scheduling is required for farm management, for organization in terms of labour planning and market sales. Harvest timing estimation in mango production is currently achieved using accumulated growing degree days (GDD) from from early stages of flower development, non-destructive estimates of fruit dry matter content by handheld near infra-red spectrometry, and destructive assessment of internal flesh colour. For fruit load estimation, current best practice involves manual counting of total fruit per tree. A range of technologies are becoming available that have relevance to assessment of mango crop harvest timing and fruit load forecast. Four activities were undertaken to assess relevant technologies: (i) A hardware system based on LoRa connected temperature sensors was characterised and recommended for field use based on measurement accuracy, battery life and reception range. An alternative algorithm on GDD calculation involving use of a function that penalises high temperatures as well as low temperatures was demonstrated to better predict harvest maturity in warmer climates. Required heat units (GDD, Tb = 12 °C, TB =32 °C) to achieve maturity were documented as 2185, 1728 and 1740 for the cultivars Keitt, Calypso, and Honey Gold, respectively. (ii) Vis-NIR spectrometry was trialled for non-invasive assessment of fruit flesh colour in the context of harvest maturity estimation, using a data set of 2034 spectra from 19 populations, where a population is an orchard/season/flowering event. The best leave-one-out-population cross validation prediction result was obtained using a Support Vector Regression (R2 of 0.63 and RMSEP of 5.52 on CIE B). However, this performance was inadequate for recommendation for use in non-invasive assessment of fruit maturity, which requires estimation to within 2.0 CIE B units. (iii) A procedure for prediction of fruit size at harvest based on measurements made prior to harvest was established, based a linear growth model for weight increment. The procedure was demonstrated for Honey Gold, Calypso and Keitt populations, with estimation error of 8.64 ± 13.7% and 0.61 ± 4.7% for measurements made between either five and four, or four and three weeks before harvest, respectively. (iv) A procedure for use of in-field machine vision-based count of fruit on tree in estimation of orchard fruit load was established, based on use of imaging on two dates to capture fruit arising from different flowering events. The two imaging estimations were accurate estimates of total orchard fruit load as measured by packhouse count, with R2 of 0.98 and slope of 0.99 across six orchards. These four activities demonstrate the potential of new technologies for improved estimation of harvest timing and load.农场管理、用工规划与市场销售环节均需开展芒果采收量与采收调度的前瞻性估算。当前芒果生产中的采收期估算，主要通过以下方式实现：基于花发育早期的累积生长度日（accumulated growing degree days, GDD）、手持式近红外光谱法无损估算果实干物质含量，以及内部果肉颜色的破坏性评估。而果实负载量估算方面，当前最优实践为人工逐树统计总果实数。目前已有一系列相关技术可用于芒果作物采收期评估与果实负载量预测。本研究开展四项活动以评估相关技术：（i）基于LoRa联网温度传感器的硬件系统：通过表征其测量精度、电池续航与接收距离，推荐该系统用于田间部署。提出一种改进的GDD计算算法：该算法采用对高温与低温均施加惩罚的函数，在温暖气候条件下可更精准地预测采收成熟期。Keitt、Calypso与Honey Gold三个品种达到成熟所需的积温单位（GDD，基准下限温度Tb=12℃、上限温度TB=32℃）分别为2185、1728与1740。（ii）可见-近红外（Vis-NIR）光谱法：针对采收成熟期估算场景，开展果肉颜色无创评估试验，共使用来自19个批次的2034条光谱数据；其中，批次指某一果园/季次/开花事件的果实样本。采用支持向量回归（Support Vector Regression）模型得到最优的留群交叉验证预测结果：在CIE B通道上，决定系数R²为0.63，均方根预测误差（RMSEP）为5.52。但该性能尚未达到可推荐用于果实成熟度无创评估的标准——此类应用要求估算误差控制在2.0个CIE B单位以内。（iii）基于采收前测量的果实采收尺寸预测流程：以果实重量增长的线性生长模型为基础，构建了采收时果实尺寸的预测方法。该流程已在Honey Gold、Calypso与Keitt三个品种的批次中得到验证：分别在采收前5至4周、4至3周开展测量时，预测误差分别为8.64±13.7%与0.61±4.7%。（iv）基于田间机器视觉的果树果实计数流程：为估算果园果实负载量，构建了田间机器视觉计数方案；该方案通过两个日期的成像，以捕获不同开花事件产生的果实。通过该流程得到的两次成像估算结果，与包装厂统计的果园总果实负载量高度吻合：在6个果园中，决定系数R²为0.98，拟合斜率为0.99。上述四项活动证实，新兴技术可有效提升芒果采收期与果实负载量的估算精度。